Target object detection method and device, storage medium and electronic device

Abstract

The application discloses a target object detection method and device, a storage medium and an electronic device, and relates to the field of computer vision and wild animal monitoring. The method comprises the following steps: collecting positioning data and environment data synchronized with an image frame; mapping the positioning data and the environment data into a conditional vector through a multilayer perception machine; linearly modulating a target convolution layer of a convolutional neural network and a target network layer of a pyramid network in a feature extraction process through the conditional vector to obtain a conditioned feature; generating a plurality of candidate image frames on the image frame according to the conditioned feature, and screening a target image frame from the plurality of candidate image frames according to the size and proportion of each candidate image frame; and detecting feature information related to a target object in the image frame according to the target image frame. The application solves the problem of insufficient detection precision and robustness of the prior art under small sample conditions for target objects with complex backgrounds and high species similarity, such as long-armed apes in tropical rainforests.

Description

Technical Field

[0001] This application relates to the fields of computer vision and wildlife monitoring, and more specifically, to a method, apparatus, storage medium, and electronic device for detecting target objects. Background Technology

[0002] Existing general object detection methods achieve acceptable accuracy when there is large-scale, high-quality labeled data, but they are prone to overfitting in small sample scenarios and their robustness to complex backgrounds decreases significantly. When the target object is highly similar to the background or similar species in color, texture, and shape, the model often misclassifies the background area or similar species as the target, leading to an increase in false positive rate and false negative rate, resulting in insufficient overall detection accuracy and robustness.

[0003] Taking the critically endangered gibbon as an example, gibbons are rare and inhabit the rainforest canopy, making it extremely difficult to obtain clear, labeled image samples. Camera traps and patrol photography commonly used in nature reserves are affected by uneven lighting, haze, foliage obstruction, and small-scale effects at long distances, resulting in images that are often small in size, low in contrast, sparse, and of inconsistent quality. Furthermore, gibbons are highly similar to closely related apes in terms of body color, size, and posture, further exacerbating false positives and false negatives. Existing transfer learning and small-sample detection methods largely rely on large-scale pre-training of basic classes, making it difficult to directly adapt to the background noise and spatiotemporal distribution biases of rainforest scenes. While contrastive learning can improve feature discriminative power, it is mostly used for classification or self-supervised tasks, lacking practical application for detection tasks, especially small-sample contrast enhancement at the candidate bounding box / region granularity. Historical sighting locations, times, and environmental elements known to nature reserves are usually only used for camera deployment or post-event manual screening, and an effective automatic discrimination mechanism has not yet been formed, resulting in insufficient detection accuracy and robustness.

[0004] There is currently no effective solution to the above problems. Summary of the Invention

[0005] This application provides a target object detection method, apparatus, storage medium, and electronic device to at least solve the technical problem in the prior art of insufficient detection accuracy and robustness for target objects with complex backgrounds and high species similarity, such as gibbons in tropical rainforests, under small sample conditions.

[0006] According to one aspect of the embodiments of this application, a target object detection method is provided, comprising: acquiring location data and environmental data synchronized with an image frame, wherein the image frame is an image frame acquired in the living environment of the target object; mapping the location data and environmental data into a conditional vector through a multilayer perceptron; using the conditional vector to perform channel-wise linear modulation on the target convolutional layer of a convolutional neural network and the target network layer of a pyramid network during the feature extraction process to obtain conditional features, wherein the target convolutional layer is used to extract image features and the target network layer is used to generate multi-scale feature maps; generating multiple candidate image boxes on the image frame based on the conditional features, and then selecting a target image box from the multiple candidate image boxes based on the size and proportion of each candidate image box; and detecting feature information related to the target object in the image frame based on the target image box.

[0007] Optionally, detecting feature information related to the target object in the image frame based on the target image box includes: drawing the target image box to obtain the region of interest in the target image box; extracting image features in the region of interest as a first image feature; proportionally expanding the first image feature based on the side length of the target image box to obtain a second image feature; fusing the first image feature and the second image feature into a target image feature; and detecting feature information related to the target object in the image frame based on the target image feature.

[0008] Optionally, fusing the first image features and the second image features into a target image feature includes: obtaining a global convergence vector between the first image features and the second image features, wherein the global convergence vector is used to compress the first image features and the second image features as multi-dimensional spatial features into a fixed-length vector; inputting the global convergence vector into a gated neural network to obtain fusion weights between the first image features and the second image features determined by the gated neural network based on contextual attention and frequency domain noise suppression mechanisms; and fusing the first image features and the second image features into a target image feature based on the fusion weights.

[0009] Optionally, detecting feature information related to the target object in the image frame based on the target image features includes: converting the target image features into multi-dimensional unit embedding vectors; performing a species similarity suppression operation based on the unit embedding vectors to obtain a species similarity suppression result, wherein the species similarity suppression operation is used to enhance the discriminative power between the target object features and the features of closely related species through metric learning and interval classification strategies, so as to form a classification boundary in the embedding space; and detecting feature information related to the target object in the image frame based on the conditional features, the species similarity suppression result, and the target image features.

[0010] Optionally, based on conditional features, species similarity suppression results, and target image features, feature information related to the target object in the image frame is detected, including: setting contrastive learning loss functions for the embedded feature information in the target image box and the embedded feature information in the region of interest, respectively, based on the target image features; and detecting feature information related to the target object in the image frame based on conditional features, species similarity suppression results, target image features, and contrastive learning loss functions set for the embedded feature information in the target image box and the embedded feature information in the region of interest, respectively.

[0011] Optionally, the first image features and the second image features are fused into target image features based on the fusion weights, including: performing a weighted calculation on the first image features and the second image features based on the fusion weights to obtain a third image feature; performing a gating operation on the third image feature to obtain the target image feature, wherein the gating operation is used to suppress feature information in the third image feature that is below a preset frequency, and retain feature information in the third image feature that is not below a preset frequency.

[0012] Optionally, the species similarity suppression operation includes: maintaining momentum prototype vectors of the target object and closely related species online, wherein the momentum prototype vector is a vector representing the features of the target object maintained through a momentum update mechanism, used to reflect the statistical characteristics of the target object features as they change over time during metric learning; calculating the cosine similarity between the unit embedding vector and the momentum prototype vector; and enhancing the discriminative power of the target object features and closely related species features based on the cosine similarity and an interval classification strategy to form a classification boundary in the embedding space, wherein the interval classification strategy is used to enhance the discriminative power of image feature representations by expanding the decision boundary between different categories in the embedding space.

[0013] Optionally, based on the target image features, contrastive learning loss functions are set for the embedded feature information in the target image bounding box and the embedded feature information in the region of interest, respectively. This includes: setting a region-level contrastive learning loss function corresponding to the embedded feature information in the region of interest, based on the target image features and the embedded feature information in the region of interest, wherein the region-level contrastive learning loss function is used to expand the inter-class margin and compress the intra-class variance at the candidate region scale; and setting an image-level contrastive learning loss function corresponding to the embedded feature information in the target image bounding box, based on the target image features and the embedded feature information in the target image bounding box, wherein the image-level contrastive learning loss function uses paired images within the same geofence and the same time window as positive samples, and images from different geofences / time windows or pure background images as negative samples.

[0014] Optionally, the target object detection method further includes: constructing a target contrast learning loss function based on a region-level contrast learning loss function and an image-level contrast learning loss function through a multi-task learning mechanism; and constructing a total loss function based on the target contrast learning loss function, the loss function set during the execution of species similarity suppression operations, and the classification loss function and regression loss function set for the target object detection task.

[0015] Optionally, the positive samples corresponding to the region-level contrastive learning loss function are multiple first-class regions of interest from the same real target, and the negative sample pairs corresponding to the region-level contrastive learning loss function are multiple regions of interest from different species or from second-class regions of interest. The crossover ratio of the first-class regions of interest is greater than a first threshold, and the crossover ratio of the second-class regions of interest is less than a second threshold, wherein the first threshold is greater than the second threshold.

[0016] Optionally, conditional features are obtained by performing channel-wise linear modulation on the target convolutional layer of the convolutional neural network and the target network layer of the pyramid network during feature extraction using conditional vectors. This includes generating channel scaling parameters and bias parameters using conditional vectors. The channel scaling parameters are learnable parameters used to dynamically adjust the response intensity of each channel in the feature map of the convolutional neural network, and the bias parameters are learnable parameters used to offset the response value of the feature map of the convolutional network channel-wise. The channel scaling parameters and bias parameters are then used to perform channel-wise linear modulation on the target convolutional layer of the convolutional neural network and the target network layer of the feature pyramid network to obtain conditional features.

[0017] Optionally, multiple candidate image boxes are generated on the image frame based on conditional features, and then a target image box is obtained by filtering from the multiple candidate image boxes based on the size and proportion of each candidate image box. This includes: using conditional features as input, generating multiple candidate image boxes at multiple scales based on the image frame through a region proposal network; and obtaining a target image box by filtering from the multiple candidate image boxes based on the size and proportion of each candidate image box, wherein the size and proportion of the target image box match the size and proportion of the target object.

[0018] According to another aspect of the embodiments of this application, a target object detection device is also provided, comprising: an acquisition unit for acquiring positioning data and environmental data synchronized with an image frame, wherein the image frame is an image frame acquired in the living environment of the target object; a first processing unit for mapping the positioning data and environmental data into conditional vectors using a multilayer perceptron; a second processing unit for using the conditional vectors to perform channel-wise linear modulation on the target convolutional layer of a convolutional neural network and the target network layer of a pyramid network during feature extraction to obtain conditional features, wherein the target convolutional layer is used to extract image features and the target network layer is used to generate multi-scale feature maps; a third processing unit for generating multiple candidate image boxes on the image frame based on the conditional features, and then selecting the target image box from the multiple candidate image boxes based on the size and proportion of each candidate image box; and a detection unit for detecting feature information related to the target object in the image frame based on the target image box.

[0019] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, which stores a computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located performs the above-described target object detection method.

[0020] According to another aspect of the embodiments of this application, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by one or more processors, the one or more processors cause the one or more processors to perform the target object detection method described above.

[0021] According to another aspect of the embodiments of this application, a computer program product is also provided, including a computer program or instructions, which are implemented when executed by a processor. According to another aspect of the embodiments of this application, a target object detection method is also provided.

[0022] In this application, the edge detection system first collects location data and environmental data synchronized with image frames, where the image frames are those acquired in the environment where the target object lives. The location data and environmental data are mapped into conditional vectors using a multilayer perceptron. Furthermore, the conditional vectors are used to perform channel-wise linear modulation on the target convolutional layer of the convolutional neural network and the target network layer of the pyramid network during feature extraction, resulting in conditional features. The target convolutional layer is used to extract image features, and the target network layer is used to generate multi-scale feature maps. Then, multiple candidate image boxes are generated on the image frame based on the conditional features. The target image box is then selected from these candidate image boxes based on their size and proportion. Finally, feature information related to the target object in the image frame is detected based on the target image box.

[0023] As described above, the edge monitoring system collects location and environmental data synchronized with image frames and encodes these data into conditional vectors using a multilayer perceptron. This data is then linearly modulated channel-by-channel using a convolutional neural network and a pyramid network. This facilitates adaptive feature extraction for geographic environment perception, transforming prior knowledge of when, where, and what environment to be considered from external filtering rules into dynamic control signals within the network. This enables the system to possess scene perception capabilities during feature extraction. For example, in historically high-incidence areas or specific environments, the system can adaptively enhance its sensitivity to weak target features; in atypical scenarios, it reduces the tendency for false alarms. By addressing the issue at the feature source, the system's robustness to complex and variable backgrounds can be improved, and targeted modulation of the feature space can partially compensate for the insufficient coverage of small sample data.

[0024] Furthermore, the edge detection system generates candidate boxes and filters target image boxes based on conditional features, and finally performs fine detection. This is equivalent to adding a morphological filter at the front end of the process, which can efficiently eliminate a large amount of irrelevant background. Under the constraint of small sample size, morphological prior screening can improve the nomination efficiency and quality, and the background suppression module overcomes the interference of complex backgrounds such as branches, leaves, and haze. In the feature space, it can forcibly separate the distinction boundary between the target object and similar objects, so that the system can still achieve high-precision and low-false-detection robust detection of gibbons under the extreme condition of small sample size. This solves the technical problem of insufficient detection accuracy and robustness of existing technologies for target objects with complex backgrounds and high species similarity, such as gibbons in tropical rainforests, under small sample size conditions. Attached Figure Description

[0025] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0026] Figure 1 This is a schematic diagram of an optional target object detection method according to an embodiment of this application;

[0027] Figure 2 This is a schematic diagram of an optional target object detection method according to an embodiment of this application;

[0028] Figure 3 This is a flowchart of an optional target object detection method according to an embodiment of this application;

[0029] Figure 4 This is a schematic diagram of the architecture and multi-task learning process of an optional target object detection method according to an embodiment of this application during the training phase.

[0030] Figure 5This is an internal processing flowchart of an optional complex background robust unit according to an embodiment of this application;

[0031] Figure 6 This is a schematic diagram of the processing flow of an optional species similarity suppression module according to an embodiment of this application;

[0032] Figure 7 This is a schematic diagram of the structure of an optional contrastive learning branch according to an embodiment of this application;

[0033] Figure 8 This is a schematic diagram of the system deployment and hardware architecture of an optional edge monitoring system according to an embodiment of this application;

[0034] Figure 9 This is a schematic diagram of an optional target object detection device according to an embodiment of this application. Detailed Implementation

[0035] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0036] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0037] According to an embodiment of this application, a method embodiment for a target object detection method is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0038] It should be noted that the information collected in this application (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for display, data used for analysis, etc.) are information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of this data all comply with relevant laws, regulations, and standards, necessary confidentiality measures have been taken, and they do not violate public order and good morals. Corresponding access points are provided for users to choose to authorize or refuse. For example, interfaces are set up between this system and relevant users or organizations, providing users with corresponding access points to choose to agree to or refuse automated decision-making results; if the user chooses to refuse, the process proceeds to the expert decision-making stage.

[0039] According to the embodiments of this application, an edge monitoring system can be used as the execution subject of the target object detection method of this application embodiment. The system can be a software system or an embedded system combining software and hardware. Of course, the method execution subject in the embodiments of this application can also be other forms of execution subject, such as devices, equipment, etc. It should be known by those skilled in the art that this application does not particularly limit the specific form of the method execution subject.

[0040] Figure 1 This is a target object detection method according to an embodiment of this application, such as... Figure 1 As shown, the method includes the following steps:

[0041] Step S101: Collect positioning data and environmental data synchronized with the image frame, wherein the image frame is an image frame collected in the living environment of the target object.

[0042] Optionally, the target species can be, but is not limited to, endangered primates such as gibbons or large mammals. Image frames can refer to image features captured in real-time by camera terminals deployed within the target species' habitat, such as RGB color images. Camera terminals can be, but are not limited to, fixed locations such as camera traps, or mobile carriers such as drones and forest mobile platforms. The camera terminals can be purely visual terminals with a resolution between 720p and 1080p, and the image content can be directly derived from the target species' actual habitat. For example, several camera terminals can be deployed within the target species' habitat based on historical location and migration corridors, combined with terrain, canopy density, and patrol routes. Each camera terminal can include: a camera unit, a satellite positioning module, an environmental data collection unit, a local computing unit, a local storage and wireless communication module, and a power supply module. The local computing unit can be an embedded AI computing chip such as an embedded graphics, tensor, or central processing unit on-chip. Image frames can be acquired by the camera terminals by collecting RGB images at a preset frame rate. It is recommended to use 720p to 1080p, which can be adapted according to computing power. At night, it can be linked to supplemental lighting or long exposure.

[0043] Optionally, after image frame acquisition, preprocessing operations such as dehazing, exposure / color temperature correction, normalization, and time / position synchronization can be performed through the preprocessing module. For example, firstly, dehazing can be performed using dark channel prior or the lightweight Retinex algorithm to eliminate atmospheric scattering caused by the high humidity environment of the rainforest; then, exposure and color temperature are corrected using the gray-world hypothesis or histogram equalization to reduce uneven lighting at dawn and dusk; next, pixel normalization and channel standardization are performed to ensure that the image numerical distribution meets the model input requirements; finally, based on the timestamp, the image is aligned with the synchronously acquired positioning data L, environmental data E, and time t to output the normalized image Î and the corresponding original side channel information (L, E, t), which facilitates subsequent conditional feature extraction.

[0044] Optionally, the location data may include latitude and longitude, altitude, and gridded location markers derived from latitude, longitude, and altitude obtained through a satellite positioning module. Environmental data may include illuminance (light intensity), meteorological conditions (such as temperature and humidity), and time periods (such as day / night, season) determined by timestamps, as well as prior geographical information such as canopy density and terrain type, and historical heat signatures. Location and environmental data synchronized with image frames are used to ensure that each image frame is precisely bound to the specific location and environmental state at the moment of acquisition through a unified timestamp, forming a multimodal synchronized data stream, providing accurate spatiotemporal and environmental context for subsequent feature-conditional modulation.

[0045] Optionally, the image frame refers to an image frame collected in the living environment of the target object to define the specific scene of image acquisition. That is, the image frame does not refer to any image in general, but specifically refers to real scene images captured by monitoring equipment deployed on historical hot spots or migration corridors in the complex rainforest environment where the target object actually moves and lives. It may include backgrounds such as branches and leaves blocking, changes in light, and haze interference. Furthermore, the target objects in the images are often small in size and have low contrast, with high realism and scene specificity.

[0046] By collecting location and environmental data synchronized with image frames, it is beneficial to ensure that the model receives not only isolated visual images, but also multimodal data streams rich in spatiotemporal context information. This facilitates the subsequent integration of prior geographical knowledge into the model and enables conditional detection that finds targets in the right place and at the right time.

[0047] Step S102: The positioning data and environmental data are mapped into conditional vectors using a multilayer perceptron.

[0048] Optionally, synchronously collected and encoded positioning data and environmental data can be used as input, and a multilayer perceptron consisting of fully connected layers can be used for nonlinear transformation and feature compression to finally output a low-dimensional, dense conditional vector. ,in, The value of can range from 32 to 128. This conditional vector condenses the spatiotemporal and environmental context semantics of the current image frame and is used to perform channel-wise linear modulation on specific layers of the convolutional neural network and the feature pyramid network (FPN). This enables the network to adaptively and dynamically adjust the feature response intensity based on prior knowledge of when, where, and what environment, thereby improving the detection sensitivity and robustness in spatiotemporal environments where the probability of target occurrence is high.

[0049] For example, condition vectors Channel scale scaling parameter γ and bias parameter β can be generated in the C3 / C4 layer of the convolutional neural network and the P3 / P4 layer of the FPN. This is beneficial for injecting "location-time-environment prior" into the network, improving the sensitivity to small and dark targets in historical activity hotspots / high-incidence periods, and reducing the false alarm threshold in atypical scenarios, thereby internalizing the monitoring strategy into a feature response.

[0050] Optionally, latitude and longitude gridding, terrain type, canopy density classification, illuminance / weather, time period, and historical heat index can be encoded as sparse or dense mixed features. For example, synchronously collected and encoded location data and environmental data can be first encoded into a dimensional grid. The dense vector is then fed into a two-layer fully connected multilayer perceptron: the first layer... The dimensions are compressed to 128 and activated by ReLU (Rectified Linear Unit). The second layer further compresses the 128 dimensions to... dimension, =64, and LayerNorm (layer normalization) is added after each layer to stabilize the distribution.

[0051] Step S103: During the feature extraction process, the target convolutional layer of the convolutional neural network and the target network layer of the pyramid network are linearly modulated channel by channel using conditional vectors to obtain conditional features. The target convolutional layer is used to extract image features, and the target network layer is used to generate multi-scale feature maps.

[0052] Optionally, the condition vector output by the multilayer perceptron... As control signals, channel-level channel scaling parameters γ and bias parameters β are dynamically generated and applied to the target convolutional layers (such as C3 and C4 layers) and the target network layers (such as P3 and P4 layers) in the convolutional neural network. The target convolutional layers are responsible for extracting visual features with rich semantic information from the input image, while the target network layers are responsible for fusing image features and converting them into multi-scale feature maps to adapt to the detection of targets of different sizes. By performing channel-wise linear modulation, the edge detection system can adaptively enhance or suppress the response intensity of each channel in the feature map based on the specific geographical, temporal, and environmental context of the current image. This gives the extracted features scene-aware capabilities, improving the model's detection sensitivity for small and low-contrast targets in historical activity hotspots or high-incidence periods, while suppressing false detections in atypical scenarios. Ultimately, this results in conditional features that are strongly correlated with the spatiotemporal environment and are more targeted.

[0053] For example, when performing channel-wise linear modulation on the target convolutional layer of a convolutional neural network and the target network layer of a pyramid network, please refer to: ,in, It can generate channel scaling parameters and bias parameters for defining mappings between the target convolutional layer of a convolutional neural network and the target network layer in a pyramid network, and for features. Do Channel-by-channel processing makes the network more sensitive to weak targets in hot zones or high-incidence periods, and reduces false alarms under atypical conditions.

[0054] The target convolutional layer in the convolutional neural network can use ResNet-50 or 101 to extract feature maps from the three residual stages C3, C4, and C5. The extracted feature maps are mapped to P3, P4, and P5 (extendable to P6) through the lateral connections of the pyramid network, with each layer having a fixed 256 channels. Subsequently, a geographic environment-aware adapter modulation is applied to the P3 and P4 stages: channel-wise scaling parameters and bias parameters are generated using synchronously acquired conditional vectors, and FiLM (Feature Linear Modulation) or CondBN (Conditional Batch Normalization) modulation is applied to the 256-dimensional features. This allows the network to dynamically adjust its response based on prior knowledge of "when, where, and what environment," thereby improving the detection sensitivity of small targets and suppressing false alarms without increasing the number of sensors.

[0055] Step S104: Generate multiple candidate image boxes on the image frame based on conditional features, and then select the target image box from the multiple candidate image boxes according to the size and proportion of each candidate image box.

[0056] Optionally, on the multi-scale conditionalized feature maps that have already undergone geographic environment-aware modulation, such as layers P3, P4, and P5, a region proposal network is used to generate a large number of preliminary candidate boxes that may contain the target. The region proposal network generates multiple candidate image boxes densely on the image frame by setting anchor boxes of different scales and aspect ratios at each location in the feature map and using convolutional layers to predict whether each anchor box contains the target and its positional fine-tuning. Then, the edge detection system, according to a preset screening strategy—primarily based on the size and shape proportions of each candidate box, combined with its category score and positional fine-tuning results—uses post-processing operations such as non-maximum suppression to select target image boxes from all candidate boxes that best match the typical body shape of the target object in terms of specific size range and limb proportions, and have high confidence. These target image boxes are then used for subsequent alignment and fine-tuning detection, facilitating rapid and targeted preliminary localization from the entire image to the potential target region.

[0057] For example, during the training and inference phases, the edge detection system can pre-define anchor boxes on multi-scale feature maps from P3 to P5, according to 5 sets of side length scales {16, 32, 64, 128, 256} and 3 sets of aspect ratios {0.5, 1, 2}. Each anchor box is processed by 3×3 convolution to extract features, and then fed into the classification head and regression head respectively for foreground or background discrimination and bounding box correction. Up to 2000 high-quality candidate boxes can be retained during training, and this can be reduced to a maximum of 1000 high-quality candidate boxes during inference to balance accuracy and speed. For densely populated regions of small targets, sampling can be intensified at layer P3 or additional anchor boxes at scale 16 can be added to improve the recall rate for small individuals.

[0058] Step S105: Detect feature information related to the target object in the image frame based on the target image frame.

[0059] Optionally, after selecting the target image boxes, the edge detection system can further perform refined feature extraction, fusion, and discriminative analysis on the target image boxes to identify whether the target image boxes contain target objects and determine the precise location and category of the target objects. For example, firstly, region features within each target image box can be precisely aligned and extracted from multi-scale feature maps through region interest alignment operations. Then, the internal features of the target image box and its context expansion features are fused using a complex background robust unit, and low-frequency background interference such as branches, leaves, and haze is suppressed through a frequency domain noise suppression mechanism to enhance discriminative features such as target edges and textures. Then, the region features are projected into normalized embedding vectors through a species similarity suppression module, and compared with the prototype vectors of the target object and closely related species maintained online. Combined with a large-margin classification strategy, the distinguishability of target object features and similar species features in the feature space is enhanced. Finally, the detection head outputs the category confidence and bounding box regression coordinates based on the processed features, thereby completing the existence determination, precise location, and differentiation from closely related species of the target object in the image frame, facilitating a complete detection process from coarse candidate box selection to refined target recognition and feature discrimination.

[0060] In one optional embodiment, detecting feature information related to the target object in an image frame based on a target image bounding box includes: an edge detection system delineating the target image bounding box to obtain a region of interest within the target image bounding box, and extracting image features from the region of interest as a first image feature; proportionally expanding the first image feature based on the side length of the target image bounding box to obtain a second image feature; then fusing the first image feature and the second image feature into a target image feature, and detecting feature information related to the target object in the image frame based on the target image feature.

[0061] Optionally, the delineation operation can refer to the Region of Interest (ROI) alignment operation. This operation precisely maps target image boxes of different locations, shapes, and sizes onto the shared feature maps extracted by the backbone network. Through alignment and bilinear interpolation, it outputs a fixed-size feature grid, ensuring that the features extracted from each candidate box are spatially aligned and regular. The resulting fixed-size feature region is the Region of Interest (ROI). The fixed size can be selected as 7×7 pixels or 14×14 pixels.

[0062] Optionally, the purpose of extracting image features from the region of interest (ROI) as the first image feature is to directly obtain the local foreground features within the target image bounding box that are most likely to contain the target object. The edge detection system can further encode the ROI features output by the delineation operation through a shallow network consisting of convolutional layers, batch normalization layers, and activation functions, thereby enhancing texture, edge, and semantic information to obtain the first image feature.

[0063] Optionally, proportional expansion refers to expanding the side length of the target image bounding box by a preset ratio r, such as r=1.5, centered on the target image bounding box, to obtain a larger rectangular region. Region interest alignment is then performed on this expanded region to extract features, thereby acquiring contextual information surrounding the target and forming a second image feature. Because target objects in rainforest environments are often obscured by branches and leaves, or blended into the background such as tree trunks and vines, the texture and structure of the surrounding environment are crucial for determining the target's presence and identifying its contours. The expansion operation facilitates the incorporation of valuable contextual clues into feature analysis.

[0064] Optionally, the method for fusing the first and second image features into the target image features can be either element-wise addition or channel concatenation. Element-wise addition refers to directly adding the corresponding pixels of the two feature maps by channel, requiring the number of channels to be exactly the same as the spatial size, and can be used for residual structures and skip-connected structures. Channel concatenation refers to stacking the two feature maps into a tensor in the channel dimension, and then using 1×1 convolution to reduce the dimensionality, keeping the spatial size unchanged while increasing the number of channels, and is suitable for scenarios that require retaining complete information from two sets of features.

[0065] Optionally, the edge detection system can input the fused target image features into the Species Similarity Suppression (SSI) module and the detection head. The SSI module converts the target image features into unit embedding vectors and compares their similarity with the target object prototype vectors maintained online, while imposing large-margin classification constraints to enhance the distinguishability with closely related species. The detection head then outputs a confidence score indicating that the target image bounding box belongs to the target object category, as well as more accurate bounding box regression coordinates, enabling the edge detection system to detect and identify target object feature information from the target image bounding box.

[0066] In one optional embodiment, fusing the first image features and the second image features into a target image feature includes: an edge detection system acquiring a global convergence vector between the first image features and the second image features, wherein the global convergence vector is used to compress the first image features and the second image features as multi-dimensional spatial features into a fixed-length vector. Then, the global convergence vector is input into a gated neural network to obtain fusion weights between the first image features and the second image features determined by the gated neural network based on contextual attention and frequency domain noise suppression mechanisms; and the first image features and the second image features are fused into the target image feature based on the fusion weights.

[0067] Optionally, the global convergence vector can be generated through a global average pooling operation. For example, an edge detection system can perform average operations on the two three-dimensional tensors, the first image feature and the second image feature, in the spatial dimension, respectively, and converge all spatial location information of each feature map channel into a scalar value. The first image feature and the second image feature can be compressed into a one-dimensional, fixed-length vector, the length of which can be consistent with the number of channels. The global convergence vector can be used to represent a global, compact semantic summary of the foreground region and the context region, so as to transform multi-dimensional features containing rich spatial details into a compact representation form that is easy for gated neural networks to perform global relational analysis and decision-making. Furthermore, the global average pooling operation has a certain degree of translation invariance, which is beneficial to the robustness of the model.

[0068] Optionally, a gated neural network refers to a lightweight multilayer perceptron that can receive a global convergence vector as input. Contextual attention mechanisms are beneficial to the learning process of the gated neural network. Through training, the gated neural network can learn to analyze the semantic relationships between the current foreground and the context. For example, when contextual features indicate that the surrounding area contains high-frequency textures or dense foliage, while the foreground features are blurry, the gated neural network can determine that the target is occluded and thus tend to give higher weights to the contextual features to aid recognition. Weights can be dynamically assigned based on semantic content.

[0069] Frequency domain noise suppression mechanisms can be incorporated into network design or loss functions. For example, after obtaining preliminary fusion features, a high-pass gating operation can be applied to the complex background robust unit. Using depthwise separable convolution approximation high-pass filtering can suppress large-area low-frequency interference such as haze, shadows, and homogenized leaf regions, while retaining mid-to-high-frequency features important for target recognition, such as hair edges and contours. During training, the gated neural network receives gradient feedback from the entire network, indirectly learning that feature combinations that better preserve effective high-frequency information and suppress low-frequency noise will be rewarded. Therefore, the fusion weights ultimately learned by the gated neural network can simultaneously balance contextual semantic relevance and feature frequency domain quality. The output of the gated neural network, after passing through a sigmoid activation function, produces a scalar fusion weight α between 0 and 1.

[0070] This application's embodiments dynamically generate fusion weights through a gated neural network, which helps to overcome the limitations of fixed-ratio fusion. This allows the system to intelligently adjust its dependence on foreground and contextual information based on the specific circumstances of each image frame and each target image box. These specific circumstances may include foreground sharpness, contextual information content, and background noise type. Furthermore, integrating contextual attention and frequency domain noise reduction into the weight learning process makes fusion not merely a simple information superposition, but an optimized choice for the final detection task of identifying specific targets in complex backgrounds. This improves the discriminative power of the fused features and enhances the quality of target image features. It provides more discriminative input for subsequent species similarity suppression modules to perform accurate measurement, comparison, and classification, facilitating high-precision, low-false-detection-rate target detection in complex backgrounds.

[0071] For example, the target image bounding box can first be delineated through region of interest (ROI) alignment to obtain a fixed-scale ROI: bilinear interpolation is used to sample the P3 / P4 / P5 feature maps, outputting a 7×7×256 ROI as the first image feature. When small targets are densely packed, the size is automatically switched to 14×14×256 to preserve fine-grained textures such as hair. Subsequently, the complex background robust unit can expand the side length of the target image bounding box by a ratio r to obtain the expanded ROI, and the ROI alignment operation can be performed again to obtain the second image feature, thus completing the context expansion. The expansion ratio r can be selected from 1.2 to 1.8.

[0072] Furthermore, robust units for complex backgrounds can employ dual-stream coding: the foreground stream pairs the first image features. Applying two layers of Conv(3×3,256)-BN-ReLU yields Context stream pairs Obtained using lightweight multi-head self-attention (MHSA, h=4 heads) or nonlocal operators Subsequently, and After global average pooling, the data are concatenated and then processed by a multilayer perceptron and a sigmoid algorithm to obtain the fusion weights α, where α ∈ [0,1]. Generating target image features facilitates the implementation of the global convergence vector → gated neural network → fusion weight process.

[0073] Optionally, frequency domain high-pass gating approximates high-pass filtering with depth-separable convolution to suppress large-area low-frequency textures or fog, such as haze and leaf shadows, while preserving mid-to-high frequency edge and hair features, making the foreground more prominent and the background smoother. The gating threshold is a learnable parameter, and the final output conditional features are sent to the detection head and species similarity suppression module to perform species similarity suppression operations, thereby forming a clear classification boundary between the target object and closely related species in the embedding space and outputting the final detection result.

[0074] In one optional embodiment, detecting feature information related to the target object in an image frame based on target image features includes: an edge detection system converting the target image features into multi-dimensional unit embedding vectors, performing a species similarity suppression operation based on the unit embedding vectors to obtain a species similarity suppression result. The species similarity suppression operation is used to enhance the discriminative power between the target object features and closely related species features through metric learning and interval classification strategies, thereby forming a classification boundary in the embedding space. Based on the conditional features, the species similarity suppression result, and the target image features, feature information related to the target object in the image frame is detected.

[0075] Optionally, after obtaining finely fused target image features, the edge detection system enters the species identification stage. This species identification stage can clearly distinguish the features of the target object from those of closely related apes with highly similar appearances in the feature space. The species identification process can be achieved through a species similarity suppression module.

[0076] Optionally, converting target image features into multi-dimensional unit embedding vectors can be accomplished using a lightweight projection network. This projection network maps target image features to a higher-level, purely semantic representation space. The projection network can include fully connected layers, batch normalization layers, and activation functions. It outputs a multi-dimensional vector, denoted as d-dimensional, where d ∈ [128, 512]. Subsequently, the edge detection system normalizes this multi-dimensional vector to a magnitude of 1, resulting in a unit embedding vector, denoted as z. This unit embedding vector normalizes target features from different images, poses, and lighting conditions into a hyperspherical space for measurement and comparison. In this hyperspherical space, feature similarity can be conveniently measured using vector angles such as cosine similarity, which is beneficial for subsequent metric learning.

[0077] Optionally, species similarity suppression can be performed using a combination of metric learning and interval classification strategies. The edge detection system does not use the species similarity suppression results in isolation, but rather fuses them with key information flows such as conditional features and target image features to make a more accurate final detection judgment. The spatiotemporal environmental priors implied by the conditional features can be used to adjust the sensitivity or confidence threshold of the final decision, making a more "bold" judgment in more likely scenarios. The original or further processed target image features can be input into a classification head to generate corresponding classification scores. The edge detection system in this embodiment fuses the scores from the classification head with the species similarity suppression results from metric learning, for example, through weighted averaging or concatenation followed by a fully connected layer. Finally, the fused comprehensive score, combining appearance details, environmental priors, and highly discriminative metric information, together with the bounding box regression results, completes the final detection of the feature information indicating the presence and precise location of the target object in the image frame.

[0078] This application combines metric learning with large-margin classification to forcibly widen the distance between the target object and closely related species feature clusters at the geometric level of the feature space, establishing a clear classification boundary. This directly suppresses false detections caused by appearance similarity at the feature representation level. The final detection decision does not rely solely on an isolated, discriminative metric head, but organically combines the species similarity suppression result, the conventional classification score, and conditional features. This enables the edge monitoring system to not only identify similar species but also retain the robustness of general target detection. Furthermore, it can adapt to specific environments, thereby achieving high-precision and highly robust target object feature information detection and recognition in complex and ever-changing real-world monitoring scenarios.

[0079] For example, in a small-sample basic adaptation scenario, a small-sample training set can be constructed using 20 annotated gibbon images as the core, supplemented by 20 images of closely related primates and 20 pure background samples. During training, the lower layers of the convolutional neural network (C1 to C2) are frozen, with only layers C3 and C4, the pyramid network, and the detection head open. A 14×14 region of interest size is used, with a contrastive learning temperature τ=0.07 and a memory queue length K=4096. The expansion ratio of robust units in complex backgrounds is r=1.5, the number of attention heads is h=4, the prototype dimension of species similarity suppression (SSI) is d=256, the update momentum is m=0.9, and the angular interval is... =0.2, scale factor s=32. After parameter tuning on the validation set, the non-maximum suppression (NMS) threshold and Soft-NMS parameters were optimized, resulting in a significant decrease in false alarms and a marked reduction in near-relative confusion in complex backgrounds. This verifies the feasibility of the region of interest extraction-feature fusion-species suppression link in data-scarce scenarios.

[0080] For example, in the near-related interference enhancement scenario, the weight of the measurement loss is further increased for 5 to 10 newly added near-related monkey samples. The value was reduced to 0.7, and triplet loss was enabled. By continuously mining easily confused samples through a priority queue of hard-to-bear samples, the angular interval between gibbons and closely related species in the embedding space was significantly increased, and the discrimination boundary in low-light and low-contrast scenes was more stable. This can reflect the strengthening effect of species similarity suppression at the metric learning level.

[0081] For example, in extreme weather scenarios, rain, fog, backlight, and nighttime samples can be introduced to unfreeze the frequency domain gating weights and allow them to learn adaptively, with the gating threshold dynamically adjusted according to the scene. Simultaneously, the GeoEnv-Adapter increases the weights of the illumination dimension, enabling the model to focus on mid-to-high frequency edges even under low light or strong scattering conditions. The results show that the foreground contour remains intact, low-frequency fog is effectively suppressed, and recall is improved, validating the robustness of conditional feature modulation and context-background refinement in harsh environments.

[0082] In one optional embodiment, feature information related to the target object in the image frame is detected based on conditional features, species similarity suppression results, and target image features. This includes: the edge detection system sets contrastive learning loss functions for the embedded feature information in the target image bounding box and the embedded feature information in the region of interest, respectively, based on the target image features. Then, feature information related to the target object in the image frame is detected based on the conditional features, species similarity suppression results, target image features, and the contrastive learning loss functions set for the embedded feature information in the target image bounding box and the embedded feature information in the region of interest, respectively.

[0083] Optionally, the embedding feature information in the region of interest (ROI) refers to the embedding vector extracted from the fixed-size region features obtained after delineation and alignment. The edge detection system can construct positive and negative sample pairs for these region-level embeddings. Positive sample pairs can come from multiple ROIs generated by different data-augmented versions of the same real target object, such as slight translations, scaling, or color jitter, requiring an Intersection over Union (IoU) greater than a high threshold, such as 0.5. Negative sample pairs include ROIs from different species, or pure background regions with extremely low IoUs to any real target, such as IoU less than 0.3. By applying contrastive loss functions such as NT-Xent loss (normalized temperature-scale cross-entropy loss), embeddings from multiple region views of the same target can be brought closer together in the feature space, compressing intra-class variance, while region embeddings from different targets or backgrounds are pushed further apart, expanding inter-class margins, thereby learning more discriminative local feature representations at the candidate region scale.

[0084] Optionally, the embedding feature information in the target image bounding box can more accurately correspond to the embedding vector extracted from the global features of the entire image frame, which can include the contextual information of the entire scene. The edge detection system constructs positive and negative sample pairs for image-level embeddings. Positive sample pairs are global embeddings of two images acquired within the same geofence and at the same time period. Negative sample pairs are global embeddings of images acquired from different geofences and at different time periods, or of pure background images that explicitly do not contain any animals. The purpose of applying the contrastive loss function is to enable the model to learn that scenes appearing at the same time and place are more likely to share the semantics of the presence of a target, and to distinguish them from scenes in different times and spaces or pure background scenes, thereby learning a robust global scene representation sensitive to the presence of a target at the entire image scale.

[0085] Optionally, conditional features can provide geographic priors, guiding the network to focus on features within a specific spatiotemporal context. Species similarity suppression results can be used through corresponding loss functions, such as spaced ArcFace loss, to force the model to learn features that clearly distinguish similar species. Target image features can serve as the basic input for all discrimination processes.

[0086] Optionally, during the inference phase, the contrastive learning branch does not participate in forward computation, thus not introducing additional latency. Based on the trained parameters, the model detects feature information related to the target object in image frames solely through the forward propagation path. However, by minimizing the total loss during training, the model can be taught to acquire highly discriminative feature representations using both region-level and image-level contrastive losses. Therefore, during inference, the model can more accurately map target image features to the correct category, i.e., the target object, and more effectively eliminate interference from background and closely related species.

[0087] This application's embodiments, by introducing region-level contrast loss, enable the model to better distinguish between foreground and background, and between target objects and similar objects at the local candidate box level, thereby improving the quality of candidate region features. By introducing image-level contrast loss, the model can learn to judge the probability of target appearance from a global scene understanding perspective, enhancing the model's context awareness and generalization ability.

[0088] In one optional embodiment, the first image features and the second image features are fused into target image features based on fusion weights. This includes: the edge detection system performing a weighted calculation on the first image features and the second image features based on the fusion weights to obtain a third image feature. Then, a gating operation is performed on the third image feature to obtain the target image feature, wherein the gating operation is used to suppress feature information in the third image feature that is below a preset frequency, and retain feature information in the third image feature that is not below a preset frequency.

[0089] Alternatively, the edge detection system can utilize the fusion weights generated by the gated neural network to optimize the first image features. Second image features Perform element-wise weighted summation: . Used to represent third image features; Used to represent the first image feature; Used to represent second image features; This is used to represent the fusion weights between the first and second image features. The first and second image features retain their original multidimensional spatial structure, and the fusion weights... The computation is broadcast to all spatial locations and channels. The target image features can be an enhanced feature representation that integrates foreground details and contextual cues, and whose weights implicitly account for robustness requirements. Element-wise weighted summation can initially mix foreground details and contextual auxiliary information in the optimal proportion for the current scene, forming an intermediate feature representation that contains richer cues but may also contain more background noise, which comes from the large-area uniform texture of the expanded context region.

[0090] Optionally, gating is used to suppress feature information in the third image features that is below a preset frequency, while retaining feature information in the third image features that is not below the preset frequency. Gating can refer to frequency domain noise suppression or high-pass gating in a Complex Background Robust Unit (CBRM), and can be implemented using a depthwise separable convolutional layer. This convolutional layer functions similarly to a learnable high-pass filter. In image processing, high and low frequencies correspond to the degree of image change: low-frequency information represents large smooth areas, uniform color patches, or slowly changing shadows, such as haze, blurred foliage in the distance, or gradual changes in lighting; low-frequency information is often background interference that needs to be suppressed. High-frequency information, on the other hand, represents sharp edges, clear textures, target outlines, and hair details, which are beneficial for identifying target objects.

[0091] Optionally, the gating operation involves passing the third image features through a depthwise separable convolutional layer. The weights of this convolutional layer are optimized during training. The training objective of these weights is to effectively filter out or significantly attenuate responses in the feature map that correspond to large areas of low-frequency background patterns (i.e., suppress feature information below a preset frequency), while allowing or even enhancing responses corresponding to high-frequency details such as edges and contours (i.e., retaining feature information at or above the preset frequency). The preset frequency is not a fixed value but a threshold implicitly defined by the filtering characteristics adaptively learned by the network under training data. The features purified by this gating operation become the target image features. The purpose of gating is to actively remove low-frequency noise introduced by complex backgrounds from the fused features, sharpen and highlight the discriminative high-frequency details related to the target, thereby improving the signal-to-noise ratio of the features and enabling the subsequent classifier to identify the target more clearly and stably.

[0092] Figure 5 A flowchart illustrating the internal processing of a robust unit against complex backgrounds is shown, which can fuse features from different sources and suppress background noise. The process begins with the first image feature. With second image features The input. Among them, the second image features... First, the features are processed by an attention / nonlocal unit to extract key contextual information. Then, the information from the first and second image features is converged into a global convergence vector, which is input into a gated multilayer perceptron. The gated neural network determines the fusion weight between the first and second image features based on contextual attention and frequency domain denoising mechanisms, outputting a fusion weight α between 0 and 1. Correspondingly, another weight is 1-α. The edge detection system performs weighted fusion of the first and second image features based on this fusion weight. The fused third image feature is then fed into a frequency domain denoising unit. This unit uses high-pass gating to suppress large areas of low-frequency background interference such as haze and uniform shadows, while retaining mid-to-high-frequency details such as edges and textures that are important for target recognition. The purified features are the final output, i.e., the target image features. .

[0093] In one optional embodiment, the species similarity suppression operation includes: an edge monitoring system maintaining online momentum prototype vectors of the target object and closely related species, wherein the momentum prototype vector is a vector representing the features of the target object maintained through a momentum update mechanism, used to reflect the statistical characteristics of the target object features as they change over time during metric learning; then, the cosine similarity between the unit embedding vector and the momentum prototype vector is calculated, and the discriminativeness between the target object features and the features of closely related species is enhanced based on the cosine similarity and an interval classification strategy to form a classification boundary in the embedding space, wherein the interval classification strategy is used to enhance the discriminativeness of image feature representations by expanding the decision boundary between different categories in the embedding space.

[0094] Optionally, the momentum prototype vector is used to characterize the dynamic vector representing the distribution center of each category's features in the embedding space. The momentum prototype vector can be denoted as: ,in, Used to represent category indexes, such as =1 can be used to represent target objects such as gibbons. The edge detection system can maintain a momentum prototype for the target object and each similar object that needs to be distinguished. The momentum prototype used for online maintenance is continuously updated throughout the training process, rather than remaining fixed, and the updates follow the momentum update mechanism.

[0095] Optionally, during metric learning, the edge detection system can maintain a vector called the momentum prototype vector for the "target object" and each closely related species that need to be distinguished. The category center. This momentum prototype vector is maintained online via a momentum update mechanism: ,in, Used to represent the momentum coefficient. ∈[0.7,0.99], for example, 0.9 is an option. Used to indicate the category in the current batch. The mean of all unit embedding vectors allows the prototype to smoothly track the evolution of class features, reflecting statistical properties. Marginal classification strategies can introduce a mandatory geometric margin during classification decisions.

[0096] For example, an edge detection system can calculate the current unit embedding vector. Momentum prototype vectors of various categories cosine similarity Alternatively, ArcFace can take the form of: Add angle intervals to real-world classes Other categories are s∈[16,64], Among them, the true category of the sample The edge monitoring system is not at the original angle Instead of directly calculating the score, it uses the true category of the sample. Add a preset angle interval , For example, it can be selected as 0.2 radians, and then calculated. For other non-realistic categories, the following method still applies. Finally, multiply all cosine values ​​by a scaling factor. , ∈[16,64], for example, The option to select 32 yields the final category score, i.e., the classification head in the form of ArcFace / CosFace. The interval classification strategy enforces a larger inter-class angular margin in the loss function, forcing the model to learn an embedding space that not only clusters similar samples around the prototype but also establishes a clear separation zone between decision boundaries of different classes, such as similar species. This enhances the discriminative power between the target object's features and those of closely related species. The distribution of classification scores output after implementing the interval classification strategy can represent the species similarity suppression result, reflecting the category confidence after a strong discriminative measure. Furthermore, triplet loss can further compress intra-class variance and strengthen the metric learning effect; for example, in semi-hard sample mining, a margin of 0.3 facilitates the implementation of species similarity suppression operations.

[0097] The momentum prototype vector maintained through the momentum update mechanism in this application provides the model with stable and progressively evolving class centers. This ensures that metric learning is based on solid evidence during training with small samples and potentially unstable feature distributions, enabling stable and adaptive class references. The margin classification strategy, by introducing a hard angular margin, does not simply penalize misclassification in the loss function, but directly imposes constraints on the geometric structure of the feature embedding space. It forces the feature distribution learned by the model to satisfy the following: not only must the centers of sample clusters of different classes be separated, but there must also be sufficiently wide classification boundaries between them, thus enforcing separability at a geometric level.

[0098] Figure 6 A schematic diagram of the processing flow of a species similarity suppression module is shown. The species similarity suppression module can receive target image features from the output of a complex background robust unit. The target image features are limited to those within the Region of Interest (RoI) to ensure accuracy. After inputting the target image features, the species similarity suppression module transforms them into multi-dimensional unit embedding vectors through an embedding and L2 normalization layer. The embedding layer represents the projection into embedding vectors, while L2 normalization ensures the target image features become unit vectors. The species similarity suppression module calculates the cosine similarity between the unit embedding vectors and the prototype vectors of each category in the prototype pool using a large-margin classification head. It also introduces an additional angular margin based on a margin classification strategy to enhance the discriminative power between the target object features and those of closely related species, forming a classification boundary in the embedding space to improve discriminativity. The edge detection system maintains an online prototype pool, storing momentum prototype vectors representing the target category and the feature centers of closely related species. This classification boundary result is fused with the output of a regular softmax classification branch to form the final output of the species similarity suppression module, thus enhancing the classification information for species discrimination.

[0099] In one optional embodiment, based on the target image features, contrastive learning loss functions are set for the embedded feature information in the target image bounding box and the embedded feature information in the region of interest, respectively. This includes: the edge detection system sets a region-level contrastive learning loss function corresponding to the embedded feature information in the region of interest based on the target image features and the embedded feature information in the region of interest, wherein the region-level contrastive learning loss function is used to expand the inter-class interval and compress the intra-class variance at the candidate region scale; then, based on the target image features and the embedded feature information in the target image bounding box, an image-level contrastive learning loss function corresponding to the embedded feature information in the target image bounding box is set, wherein the image-level contrastive learning loss function uses paired images within the same geofence and the same time window as positive samples, and images from different geofences / time windows or pure background images as negative samples.

[0100] Optionally, a region-level contrastive learning loss function can be enabled during training to enhance the differentiation between the target object and closely related species and the background. Positive sample pairs can come from image features of the same real target after different jittering or scaling transformations, with an IoU ≥ 0.5; negative sample pairs can include image features of closely related apes and background images (IoU ≤ 0.3), with priority given to samples that are similar in appearance and texture. Projected onto a multilayer perceptron, it becomes a 128-dimensional vector, i.e. After normalization, normalized temperature cross-entropy or supervised contrast loss is used, with temperature τ∈[0.05,0.2]. A memory queue of length K∈[1024,8192] can be introduced to expand the diversity of negative samples, which is beneficial to realize the metric learning of species similarity suppression at the regional level.

[0101] Optionally, image-level contrast focuses on scene-level target presence. The feature maps from the convolutional neural network are processed through global average pooling followed by a fully connected layer to obtain a 128-dimensional unit vector h. Positive sample pairs can be two views from the same geofence, within the same time window, or augmented with different data; negative sample pairs can be images from different geofences, different time windows, or pure background images. The loss function is the same as that at the region level. The global representation learned through image-level contrast can compensate for situations where local image features fail to fully represent the target, working synergistically with the region-level approach to further improve the robustness and generalization ability of gibbon detection under small sample conditions.

[0102] Optionally, the embedded feature information in the region of interest (ROI) refers to the feature vector extracted from the aligned first image features or after further projection of the first image features. The edge detection system constructs positive and negative sample pairs for contrastive learning based on the region-level embedding of the ROI. Positive sample pairs can originate from the same real target object. Multiple different ROI views are generated by slightly jittering, scaling, or color enhancement of the target region, requiring an Intersection over Union (IoU) greater than a first threshold, such as 0.5. Negative sample pairs include ROIs from different species and background regions with an IoU less than a second threshold, which can be optionally 0.3. The edge detection system can apply contrastive losses such as normalized cross-entropy loss. A discriminative constraint at the candidate region scale can be introduced into the model's training objective. This region-level contrastive learning loss function compresses intra-class variance by bringing together multiple view embeddings of the same target while pushing apart embeddings of different targets or backgrounds. This makes different appearance variants of the same target more clustered in the feature space and expands the inter-class margin, making the target object, closely related species, and background more separated in the feature space. This can improve the model's representation quality of local regions and its ability to distinguish subtle differences.

[0103] Optionally, the embedding feature information in the target image bounding box is used to characterize the embedding vector extracted from the global features of the entire image frame, which is used to characterize the semantics of the entire scene. The edge detection system constructs contrastive sample pairs for image-level embedding. The construction of positive sample pairs depends on the deployment spatiotemporal prior: two images collected within the same geofence and within the same time window can be selected for different global enhancements. Negative sample pairs include: images collected at different geofences, images collected at different time windows, and pure background images confirmed not to contain any animals. In this embodiment, an image-level supervision signal can be introduced. The image-level contrastive learning loss function allows the model to learn that the appearance of the target object has a strong spatiotemporal correlation. It can teach the model to regard scenes in the same spatiotemporal context as semantically similar, while considering scenes with different appearance patterns as dissimilar. This helps the model learn a more generalizable global scene representation that is more sensitive to the existence of the target.

[0104] The region-level comparison in this application focuses on improving the model's fine-grained ability to identify targets within a single candidate box, which can optimize the quality of local detection in the detection process; while the image-level comparison, from the macro perspective of the entire scene, guides the model to understand the spatiotemporal patterns of target appearance, optimizes the contextual awareness ability of global detection, and can work together at different scales so that the features learned by the model have both local discriminative power and global rationality, and can achieve a synergistic improvement of local and global discriminative power.

[0105] For example, Figure 7 A schematic diagram of a contrastive learning branch is shown. The structure of the contrastive learning branch includes two parallel contrastive learning paths: region-level and image-level. In the region-level contrastive learning path, shared features extracted by the convolutional neural network (Backbone) and pyramid network (FPN) are processed by a complex background robust unit (CBRM) to obtain the first image features, which can be mapped to a region embedding vector z through a region of interest projection head. This path uses the first image features from multiple viewpoints (IoU ≥ 0.5) of the same real target as positive sample pairs, and the target image features of closely related apes as negative sample pairs, and applies the NT-Xent loss function with a temperature parameter τ to calculate the region-level contrastive learning loss. In the image-level contrastive learning path, deep global features (such as layer C5) extracted by the convolutional neural network (Backbone) and pyramid network (FPN) are mapped to image embedding vectors h through a global projection head. This path uses images within the same geofence and acquired in the same time window as positive sample pairs, and images with different geofences or pure backgrounds as negative sample pairs. The NT-Xent loss is also applied to calculate the image-level contrastive learning loss. .

[0106] In an optional embodiment, the target object detection method further includes: the edge detection system constructs a target contrast learning loss function based on a region-level contrast learning loss function and an image-level contrast learning loss function through a multi-task learning mechanism. Then, based on the target contrast learning loss function, the loss function set during the execution of the species similarity suppression operation, and the classification loss function and regression loss function set for the target object detection task, a total loss function is constructed.

[0107] Optionally, a multi-task learning mechanism refers to a strategy that simultaneously optimizes multiple related but differentiated subtasks, such as contrastive learning, classification, and regression, to achieve knowledge sharing and performance improvement. In this application, the edge detection system treats the region-level contrastive learning loss function and the image-level contrastive learning loss function as two collaborative subtasks to enhance feature discrimination. These can be combined in a weighted sum to jointly constitute the target contrastive learning loss function, i.e. ,in, and Used to represent adjustable hyperparameter weights, for example, values ​​can range from 0.2 to 1.0. Used to represent the target contrastive learning loss function; Used to represent the region-level contrastive learning loss function; Used to represent the image-level contrastive learning loss function.

[0108] Constructing a target contrastive learning loss function can integrate contrastive learning supervision signals from different granularities, such as region level and image level, to form a unified contrastive learning optimization objective aimed at comprehensively improving the quality of model feature representation. This ensures that the model can simultaneously receive supervision signals from different local views that bring the same target closer together and from different global scenes that are associated with the same spatiotemporal context during training, enabling the model to learn powerful features that are both locally discriminative and globally consistent.

[0109] Optionally, the two newly added optimization objectives, the region-level contrastive learning loss function and the image-level contrastive learning loss function, are compared with the basic detection loss and the aforementioned metric loss during training. Together, they form a multi-task learning framework. The edge detection system performs end-to-end optimization through a total loss function, which is: ,in, Used to represent the basic detection loss, , Used to represent classification loss, Used to represent regression loss; Used to represent the measurement loss Used to represent regional-level contrast loss. Used to represent image-level contrast loss , , The weighting coefficients are used to represent the weights of each loss. The target contrast learning loss function can improve feature discrimination, the loss function set during the species similarity suppression operation can enhance the ability to distinguish similar species, the classification loss function set for the target object detection task can be used to directly optimize the model's accuracy in binary or multi-class classification of whether the candidate box contains the target object, and the regression loss function is used to optimize the model's prediction accuracy of the target object's bounding box position and size. By weighting, a comprehensive and balanced optimization objective can be established, enabling the model to simultaneously optimize detection accuracy, feature discriminability, and species discrimination precision during training. This helps ensure that the model's various capabilities develop synergistically and promote each other in end-to-end training.

[0110] This application's embodiments organically integrate basic detection tasks, fine-grained species discrimination tasks, and dual-granularity feature enhancement tasks through a multi-task learning mechanism. This reduces potential sub-objective conflicts or neglects that might result from individual optimization, guiding the model to simultaneously achieve accurate localization, accurate classification, strong discrimination, and robust generalization. The total loss function can constitute a complete supervision hierarchy from bounding box regression to feature space discrimination, ensuring that all levels of the model, from low-level visual features to high-level semantic representations, are specifically guided and constrained. This allows the model to more efficiently and fully utilize limited annotation information under small sample conditions, learning highly effective inherent rules for generalization, thus forming a hierarchical loss supervision system.

[0111] Figure 4 This paper illustrates the architecture and multi-task learning process of a target object detection method during the training phase. The processing begins with an image frame and auxiliary data synchronized with the image frame, namely localization data and environmental data. The input localization data and environmental data are extracted and modulated into conditional features by a convolutional neural network and a pyramid network. These conditional features are then input into a region proposal network, which generates candidate image boxes through operations such as dimensionality transformation and the softmax function. These candidate image boxes extract image features within the regions of interest (i.e., delineation operations) to obtain feature maps. Subsequently, the image features are fed into a Complex Background Robust Unit (CBRM), which may contain a gated multilayer perceptron, attention / nonlocality, and frequency domain noise suppression components for background suppression. The processed features are then input into a Species Similarity Suppression (SSI) module, which involves unit embedding vector, momentum prototype vector update, and fusion operations to enhance species differentiation. Finally, the detection head performs localization and classification, outputting the detection results. During training, the model is jointly optimized based on multiple losses: including a basic detection loss consisting of localization and classification losses for basic detection. Measured loss generated by the species similarity suppression module Regional-level contrastive learning loss based on regional features and image-level contrastive learning loss based on global features The total loss function is: This can reflect a multi-task learning mechanism.

[0112] In one optional embodiment, the positive samples corresponding to the region-level contrastive learning loss function are multiple first-class regions of interest from the same real target, and the negative sample pairs corresponding to the region-level contrastive learning loss function are multiple regions of interest from different species or from second-class regions of interest. The cross-union ratio of the first-class regions of interest is greater than a first threshold, and the cross-union ratio of the second-class regions of interest is less than a second threshold, wherein the first threshold is greater than the second threshold.

[0113] Optionally, the first type of region of interest refers to candidate regions generated through data enhancement such as jittering and scaling, whose intersection-union ratio (IoU) with the labeled real gibbon target bounding box is greater than a first threshold, which can be greater than or equal to 0.5. The positive samples corresponding to the region-level contrastive learning loss function, being multiple first-type regions of interest from the same real target, ensure that the positive samples visually correspond strictly to the same gibbon instance. Through contrastive learning, the model brings together the feature embeddings of the same target under different viewpoints or transformations, thereby compressing intra-class variance and improving the model's robustness to changes in target appearance.

[0114] Optionally, the second type of region of interest refers to regions with an intersection-union ratio (IU) less than a second threshold that are adjacent to any real target, such as background or hard samples containing very few targets. The first threshold being greater than the second threshold can create a clear quality stratification. The negative sample pairs corresponding to the region-level contrastive learning loss function can be multiple regions of interest from different species or from the second type of region of interest. This allows for hard negative sample mining, i.e., using regions of closely related species with similar appearances as negative samples can expand the inter-class margin, forcing the model to strengthen its species discrimination ability; using background regions with low IU as negative samples can enhance the model's ability to separate foreground from complex backgrounds.

[0115] This application embodiment precisely defines positive and negative samples by setting a quantified cross-union ratio threshold, which can construct a high-quality contrastive learning supervision signal with distinct levels and strong targeting. It can significantly improve the optimization efficiency and targeting of the region-level contrastive learning loss function, and ensure that the model directly faces and overcomes the challenges of distinguishing similar species and separating complex backgrounds during training. This drives the model to learn a region-level feature representation with strong discriminative power, enabling high-precision detection under small sample conditions.

[0116] Optionally, the joint training process and hyperparameters may include:

[0117] First, data organization is crucial: due to the extreme scarcity of gibbons, at least 30 effectively labeled images are required, covering multiple poses, distances, and lighting conditions. The edge detection system can simultaneously collect negative samples from closely related gibbons such as macaques and crested gibbons, as well as pure background, which helps compensate for insufficient samples. Before training, image enhancements can be applied, such as random cropping, color jittering, contrast or brightness changes, simulated fog or raindrops, leaf texture occlusion patches, and random rotation or horizontal flipping. Random cropping retains ≥70% of the target pixels. The sampling strategy requires at least four images per batch, ensuring a balance among the target, closely related, and background classes. The positive-to-negative ratio of the region of interest is fixed at 1:3, with priority given to selecting difficult negative samples to improve discriminative power.

[0118] Furthermore, the loss function is confirmed: the detection branch adopts... ,in, You can choose either CE (cross-entropy) or Focal (focus loss). Optional SmoothL1 loss or GIoU loss can be used for localization and classification. Metric branch. ArcFace loss (additive angular margin loss) or CosFace loss (large margin cosine loss) can be used to increase the distance between individuals, and triplet loss can also be added. The contrastive branch includes a region-level contrastive learning loss function. Image-level contrastive learning loss function Each of these measures suppresses false alarms. The overall joint loss function is... Typical weights =0.5、 =0.5、 =0.3; In the early stages of training, it can be... or Set the value to a slightly lower level, and gradually increase it after the classification stabilizes, in order to balance the convergence speed and feature discriminability.

[0119] Then, staged training is carried out: (1) Base class pre-training: Faster R-CNN is pre-trained on a general dataset. At this time, the contrastive branch is not enabled, and only the general target features are learned. (2) Few-shot fine-tuning: The low layer of the convolutional neural network (stem-C2) is frozen, and C3 / C4 and the detection head are enabled. At the same time, the GeoEnv-Adapter, Complex Background Robust Unit (CBRM), Species Similarity Suppression (SSI) and contrastive branch are activated, and fine-tuning is carried out with 30 gibbons as the core. (3) Calibration: The classification threshold and non-maximum suppression (NMS / Soft-NMS) parameters are fitted on the validation set, and secondary screening is carried out in combination with geofencing and time window rules. Finally, an acceptable balance between the mean average accuracy (mAP) and the false alarm rate is achieved.

[0120] Finally, the typical hyperparameters were determined, with AdamW selected as the optimizer and an initial learning rate of 1×1. ³-5×1 A cosine annealing strategy is employed. Due to embedded memory limitations, the batch size is set to 2 to 4, with an equivalent batch size of 8 achieved through gradient accumulation. Training iterations are performed for 5k to 20k steps, with an early stopping mechanism to prevent overfitting. A contrast learning and memory queue length of K=4096 and a contrast temperature τ=0.07 are used to ensure that all training and inference can be completed at the edge.

[0121] Optionally, after training at the edge, the edge detection system enters the inference phase: only the detection convolutional neural network, the Complex Background Robust Unit (CBRM), and the Species Similarity Suppression (SSI) module are retained, and the contrastive learning branch is no longer enabled. The GeoEnv-Adapter generates conditional vectors based on real-time synchronized location data (L), environmental data (E), and timestamps (t), and performs channel-wise linear modulation on the target convolutional layer in the convolutional neural network and the pyramid network (FPN) to complete one forward inference. Subsequently, non-maximum suppression (NMS) is used, with a threshold selectable at 0.5, to remove redundant bounding boxes; if the scene occlusion is severe, it switches to soft non-maximum suppression (Soft-NMS), with a threshold selectable at 0.5, to balance recall and localization accuracy.

[0122] Optionally, after inference is completed, automatic filtering and reporting can be performed. The inference results must go through four layers of automatic filtering before they can be uploaded: 1) Geofencing: Only gibbon candidate boxes located within the preset monitoring area are retained; 2) Time window: The confidence threshold is appropriately relaxed during known peak activity periods and tightened during off-peak periods to reduce false alarms; 3) Spatiotemporal consistency: By matching the trajectories of adjacent frames or adjacent cameras, candidates with isolated single frames and low confidence are downweighted or directly discarded; 4) Reporting: The final retained bounding boxes, confidence, spatiotemporal labels and thumbnails are sent to the cloud platform via wireless link to enter the manual sampling and long-term archiving process, realizing closed-loop monitoring of "end-cloud" collaboration.

[0123] Optionally, after deployment at the edge, the edge monitoring system can achieve energy consumption and latency control through three configurable operating strategies: (1) Low-power cruise: actively reduce the resolution and frame rate of the input image, retain only the lightweight paths of the GeoEnv-Adapter and the Regional Proposal Network (RPN), and continuously monitor with an average power consumption of ≤4 W; (2) Event triggering: use motion detection or optical flow threshold to judge the changes in the image, and wake up the complete model (Backbone+CBRM+SSI) for high-precision inference once triggered; (3) High-precision window: when the target is in a historical hot zone or a known high-incidence period, the edge monitoring system automatically switches to 720p high resolution and full feature mode to ensure that the single-frame inference latency is ≤120ms, taking into account both real-time performance and detection accuracy.

[0124] Optionally, the convolutional neural network (backbone) and detector head can employ INT8 quantization (8-bit integer quantization), while the Complex Background Robust Unit (CBRM) and Species Similarity Suppression (SSI) module maintain FP16 precision (16-bit floating-point precision) to preserve critical performance and meet the computational and memory constraints of the System-on-Chip (SoC). Furthermore, through channel pruning and knowledge distillation, the original backbone is compressed into a lightweight network (such as ResNet-18 or MobileNetV3), achieving a single-frame inference time of ≤120ms and an average power consumption of <4W at 720p input, which facilitates hardware- and policy-dependent configurability.

[0125] In this embodiment, the GeoEnv-Adapter generates conditional vectors based on real-time positioning and environmental data, and performs Feature Linear Modulation (FiLM) on the C3 / C4 layers of the convolutional neural network and the P3 / P4 layers of the pyramid network (FPN). This injects prior knowledge of "when, where, and what environment" into the feature space, thereby improving the initial recall of gibbons with very few labeled samples. Secondly, the Complex Background Robust Unit (CBRM) performs context expansion and dual-stream encoding on the region of interest (RoI), using multi-head self-attention (MHSA) and frequency domain high-pass gating to suppress low-frequency interference such as fog and leaf shadows, while retaining mid-to-high-frequency information such as hair edges, reducing foreground and background confusion in complex rainforest backgrounds. Thirdly, the Species Similarity Suppression (SSI) module establishes large-margin classification boundaries in the embedding space through prototype metric learning, and further compresses intra-class variance and expands inter-class distance by combining region-level and image-level contrastive learning, significantly reducing false positives for closely related apes. Finally, the two-stage detection framework of the Region Proposal Network (RPN) and the detection head, through the coordination of quantization and event-triggered strategies, maintains a simple inference path: only the lightweight RPN runs during low-power cruise, and the full model is activated after triggering, enabling real-time edge inference with 720p single frame ≤120ms and average power consumption <4W. Thus, the synergistic effect of prior conditionation, context-frequency domain robustness, prototype metric-large-interval discrimination, dual-layer contrast enhancement, and robust backbone enables the edge monitoring system to maintain high accuracy, low false alarms, and low power consumption even under conditions of small sample sizes (no more than 30 labeled samples), complex rainforests, and near-edge interference. Furthermore, all parameters and modules can be flexibly tailored to the actual conditions of the protected area.

[0126] In one optional embodiment, conditional vectors are used to perform channel-wise linear modulation on the target convolutional layer of the convolutional neural network and the target network layer of the pyramid network during feature extraction to obtain conditional features. This includes: the edge detection system can use conditional vectors to generate channel scale scaling parameters and bias parameters; wherein, the channel scale scaling parameters are used to dynamically adjust the learnable parameters of the response intensity of each channel in the feature map of the convolutional neural network, and the bias parameters are used to offset the learnable parameters of the response values ​​of the feature map of the convolutional network channel-wise; then, through the channel scale scaling parameters and bias parameters, channel-wise linear modulation is performed on the target convolutional layer of the convolutional neural network and the target network layer of the feature pyramid network to obtain conditional features.

[0127] Optionally, the conditional vector is used to specify a low-dimensional semantic vector after the bit data and environmental data are encoded by a multilayer perceptron. The edge detection system generates two sets of parameters for this conditional vector through a specific learnable mapping function: a channel scaling parameter γ and a bias parameter β. Here, γ can be a vector, where each element is a learnable scaling factor used to dynamically adjust the response intensity of all spatial locations of the corresponding channel in the convolutional neural network feature map; this can be understood as controlling which features are more important in the current environment. β can also be a vector of equal length, where each element can be a learnable offset used to globally shift the response values ​​of the corresponding channel in the convolutional neural network feature map channel by channel; this can be understood as calibrating the feature baseline according to the environment. Using the conditional vector to generate the channel scaling parameter and bias parameter transforms the abstract environmental semantics into a concrete debugging tool that can directly and finely manipulate the feature map values.

[0128] Optionally, during feature extraction, when the network forward propagates to the preset target convolutional layer and target network layer, the edge detection system extracts the original feature map x output by the preset target convolutional layer and target network layer and performs the operation: x′=γ⊙x+β, where ⊙ represents channel-wise multiplication. For example, implementing channel-wise linear modulation on the C3 / C4 layer of the convolutional neural network and the P3 / P4 layer of the pyramid network can inject geographical environment priors into visual features, enabling the network to adaptively enhance or suppress the responses of different semantic channels based on the specific spatiotemporal context of the input image (such as whether it is foggy weather or located in a historical activity hotspot). For example, under low-light conditions, the system may enhance edge and texture-sensitive channels through γ and β, thereby obtaining more discriminative conditional features that are strongly correlated with the environment.

[0129] The embodiments of this application can dynamically control the fine-grained, adaptive, and scene-aware features extraction process, transforming external environmental knowledge into learnable transformations within the network. This improves the model's feature extraction targeting and robustness for gibbons in complex and variable rainforest environments, thereby enhancing small-sample detection performance.

[0130] In one optional embodiment, multiple candidate image boxes are generated on an image frame based on conditional features. Then, a target image box is selected from the multiple candidate image boxes based on the size and proportion of each candidate image box. This includes: an edge detection system using conditional features as input generates multiple candidate image boxes at multiple scales based on the image frame through a region proposal network. Then, a target image box is selected from the multiple candidate image boxes based on the size and proportion of each candidate image box, wherein the size and proportion of the target image box match the size and proportion of the target object.

[0131] Optionally, conditional features refer to the multi-scale feature maps obtained after geographic environment-aware modulation. The region proposal network pre-defines a series of anchor box scales and aspect ratios on the multi-scale feature maps that match the possible sizes of the target object in the image. It then predicts the confidence and positional offset of each anchor box containing the foreground through convolution, thereby densely generating a large number of candidate image boxes. Using conditional features as input, the region proposal network generates multiple candidate image boxes at multiple scales based on image frames, enabling a comprehensive scan of the image and preliminary localization of all potential regions that may contain the target object, ensuring a high recall rate.

[0132] Optionally, during the screening process, the edge detection system can prioritize retaining candidate boxes whose predicted box sizes fall within the typical range of gibbon size and whose aspect ratio matches their slender physique. Simultaneously, it can combine the confidence level of the RPN predictions for sorting and quantity control. The purpose of filtering target image boxes from multiple candidate image boxes based on their size and aspect ratio is to perform a coarse filtering based on target morphology priors, eliminating a large number of background regions that clearly do not conform to the gibbon's body shape characteristics. This reduces the number of candidate boxes to a controllable, high-quality set, i.e., the target image boxes.

[0133] The embodiments of this application can achieve rapid and highly recall region nomination and knowledge-based intelligent initial screening. While ensuring that no target is missed, it can significantly reduce the computational burden of subsequent fine processing, which is conducive to improving the operating efficiency of the entire edge monitoring system on edge devices.

[0134] For example, Figure 2A schematic diagram of a target object detection method is shown. First, an image frame is acquired by a camera terminal, simultaneously acquiring location and environmental data. The location and environmental data undergo preprocessing by a data acquisition and preprocessing unit, followed by synchronous multimodal adaptation, alignment, and fusion to provide standardized input for subsequent processing. Then, multi-scale features of the image are extracted using a convolutional neural network and a pyramid network. The feature extraction process of this network is based on channel-wise linear modulation of the conditional vector generated from the synchronous data, resulting in conditional features. These conditional features are input to a region proposal network to generate multiple candidate image boxes on the image frame, and also provide feature support for subsequent steps. The candidate image boxes undergo a region interest alignment operation, i.e., delineation, to obtain features in the region of interest within the target image box. For example, first and second image features can be extracted from the region of interest. These image features are then fed into a complex background robust unit for context fusion and frequency domain noise suppression to suppress interference from complex backgrounds. The output of the complex background robust unit is then fed into a species similarity suppression module, where metric learning and interval classification strategies enhance species discrimination. Finally, the processed features are analyzed by the detection head to detect feature information related to the target object in the image frame and output a target detection result warning.

[0135] Figure 3 A flowchart of a target object detection method is shown. First, the camera terminal performs acquisition, preprocessing, and synchronization to obtain location data and environmental data synchronized with the image frame. Then, the location data and environmental data are mapped into conditional vectors via a GeoEnv-Adapter and a multilayer perceptron. These conditional vectors are then used to perform channel-wise linear modulation on a convolutional neural network and a pyramid network to obtain conditional features. Next, based on the conditional features, a candidate region proposal network generates multiple candidate bounding boxes on the image frame. These candidate bounding boxes undergo region interest alignment and complex background robustness processing, corresponding to the RoIAlign and CBRM modules, to extract features and suppress background interference. The processed features are then fed into a species similarity suppression (SSI) module to enhance the distinguishability with closely related species. During the training phase, the edge detection system performs joint optimization through multi-task learning and contrast mechanisms, combining region-level and image-level contrastive learning losses. Finally, in the inference phase, inference and post-processing are performed to detect feature information related to the target object in the image frame.

[0136] Optionally, the edge detection system can primarily use visual input, including a data acquisition and preprocessing unit, a geographic environment perception adaptation unit, a backbone feature extraction and feature pyramid unit, a region proposal network unit, a region of interest alignment unit, a complex background robustness unit (CBRM), a species similarity suppression unit (SSI), and a detection head and post-processing unit. Location and environmental data are not used as the primary modality for recognition, but rather as conditional priors entered into the geographic environment perception adaptation unit. After acquiring images, the data acquisition and preprocessing unit can first perform dehazing and exposure / color temperature correction to ensure the stability of subsequent features. The geographic environment perception adaptation unit encodes location and environmental information into conditional vectors, and uses Feature Linear Modulation (FiLM) to perform channel-level modulation on the C3 / C4 layers of the convolutional neural network and the P3 / P4 layers of the feature pyramid network to achieve prior injection of "when, where, and what environment". After the backbone and pyramid units output multi-scale features, the region proposal network generates candidate boxes, and the region of interest alignment unit extracts regions of interest at a fixed scale. The complex background robust unit performs context expansion, dual-stream coding, and frequency domain high-pass gating on this region to suppress low-frequency interference such as fog and leaf shadows, while retaining key information such as hair edges. The species similarity suppression unit uses prototype metric learning to establish large-margin classification boundaries in the embedding space, and further compresses intra-class variance and expands inter-class distance through region-level and image-level comparisons, significantly reducing false positives for closely related apes. After the detection head outputs the category and bounding box, the post-processing unit uses Soft-NMS (σ∈[0.4,0.6]) to remove redundant boxes and combines geofencing and time window rules to complete automatic filtering, achieving high-precision, low-false-positive edge inference.

[0137] The edge monitoring system can also be equipped with an energy consumption and latency control unit to meet the power consumption and latency constraints of long-term deployment in protected areas. This control unit adaptively switches between resolution, frame rate, and quantization accuracy based on time period (high-incidence / low-incidence) and event triggering (motion detection / optical flow threshold): during low-power cruise, the resolution and frame rate are reduced, and only a lightweight region suggestion network is run; after an event trigger, it switches to full-precision inference, with the convolutional neural network and detection head using INT8 quantization, and CBRM / SSI retaining FP16 to ensure critical accuracy. The non-maximum suppression threshold can be dynamically adjusted within the range of 0.4 to 0.6 to ensure high recall even in occluded scenarios. Through gradient accumulation, channel pruning, and knowledge distillation to a lightweight backbone, the edge monitoring system achieves single-frame inference ≤120ms and average power consumption <4W at 720p input, meeting the real-time monitoring requirements at the edge. For long-term field monitoring, an integrated device solution of "automatic screening and identification—spatiotemporal filtering—alarm reporting—low-power edge deployment" can be formed.

[0138] For example, Figure 8This paper illustrates the system deployment and hardware architecture of an edge monitoring system. The edge monitoring system may include a front-end acquisition unit and an intelligent processing unit. The front-end acquisition unit includes a camera, a positioning module, and an infrared trigger sensor, responsible for data acquisition, and can be encapsulated in a waterproof casing and powered by a power supply module. The data acquired by the front end is transmitted through the front-end processing module (with a Wi-Fi module). The intelligent processing unit can be powered by solar energy and may include multiple functional modules: a data acquisition and preprocessing unit, capable of preprocessing and synchronizing image frames, positioning data, and environmental data for alignment; a policy controller, for managing operating modes; a GeoEnv conditionalization module (i.e., GeoEnv-Adapter), for performing feature modulation; a detection inference module, for running detection algorithms; a fusion and comparison module, for implementing contrastive learning optimization during training; and a result caching module, for temporarily storing results. Finally, the data can be uploaded to a cloud computing platform for further analysis and archiving.

[0139] Additionally, the conditional vector dimension K of the GeoEnv-Adapter is recommended to be 32 to 128, and it is recommended to inject it into the C3 / C4 layer and the P3 / P4 layer of the feature pyramid in the convolutional neural network. The context expansion ratio r of the Complex Background Robust Unit (CBRM) can be selected from 1.2 to 1.8, the number of heads in the Multi-Head Self-Attention (MHSA) can be selected from 2 to 6, and the depthwise separable convolutional kernel can be selected from 3×3 or 5×5. The prototype dimension d of Species Similarity Suppression (SSI) can be selected from 128 to 512, the update momentum m can be selected from 0.7 to 0.99, and the angular interval... The scale factor s can be selected from 0.1 to 0.4, and the scale factor s can be selected from 16 to 64. The prototype maintenance of species similarity suppression can be changed to "class center + Bayesian update"; the contrast learning temperature τ can be selected from 0.05 to 0.2, the memory queue length K can be selected from 1024 to 8192, and each batch should have at least 1 positive sample pair and 4 negative sample pairs. The anchor scale of the Region Proposal Network (RPN) can be fine-tuned according to the field of view of the deployed camera, the non-maximum suppression (NMS) IoU threshold can be selected from 0.45 to 0.55, and Soft-NMS can be used when occlusion is severe, with σ selected from 0.4 to 0.6.

[0140] Optionally, parameters can be extended and equivalently replaced. For example, in addition to ResNet (Residual Network), convolutional neural networks can be replaced with Swin-T transformers, ConvNeXt convolutional networks, etc. Conditioning methods, besides FiLM (Feature-wise Linear Modulation), can also employ compression-excitation (SE), gated bias, or conditional attention; metric learning heads, besides ArcFace / CosFace, can also use CenterLoss (Center Loss) or Proxy-NCA (Proxy Nearest Neighbor Component Analysis Loss); the loss function of the contrastive branch can be replaced with similar forms such as SupCon (Supervised Contrastive Loss) or InfoNCE (Information Noise Contrastive Estimation Loss). With simple adjustments to the prototype set and a small number of samples, it can be extended to the monitoring of other endangered primates or large mammals, demonstrating good versatility.

[0141] See Figure 9 According to another aspect of the embodiments of this application, a target object detection device is also provided, including: a collection unit 901, a first processing unit 902, a second processing unit 903, a third processing unit 904, and a detection unit 905.

[0142] The system includes: an acquisition unit 901, used to acquire location data and environmental data synchronized with image frames, wherein the image frames are acquired in the environment in which the target object lives; a first processing unit 902, used to map the location data and environmental data into conditional vectors using a multilayer perceptron; a second processing unit 903, used to perform channel-wise linear modulation on the target convolutional layer of the convolutional neural network and the target network layer of the pyramid network during feature extraction using the conditional vectors to obtain conditional features, wherein the target convolutional layer is used to extract image features and the target network layer is used to generate multi-scale feature maps; a third processing unit 904, used to generate multiple candidate image boxes on the image frame based on the conditional features, and then select the target image box from the multiple candidate image boxes according to the size and proportion of each candidate image box; and a detection unit 905, used to detect feature information related to the target object in the image frame based on the target image box.

[0143] Optionally, the detection unit 905 includes: a delineation operation processing subunit, used to perform a delineation operation on the target image box to obtain a region of interest in the target image box; a first image feature extraction subunit, used to extract image features in the region of interest as first image features; a second image feature extraction subunit, used to proportionally expand the first image features based on the side length of the target image box to obtain second image features; a target image feature determination subunit, used to fuse the first image features and the second image features into target image features; and a feature information detection subunit, used to detect feature information related to the target object in the image frame based on the target image features.

[0144] Optionally, the target image feature determination subunit includes: a global convergence vector acquisition module, used to acquire a global convergence vector between the first image feature and the second image feature, wherein the global convergence vector is used to compress the first image feature and the second image feature as multi-dimensional spatial features into a fixed-length vector; a fusion weight determination module, used to input the global convergence vector into a gated neural network to obtain the fusion weight between the first image feature and the second image feature determined by the gated neural network based on contextual attention and frequency domain noise reduction mechanisms; and a target image feature determination module, used to fuse the first image feature and the second image feature into a target image feature based on the fusion weight.

[0145] Optionally, the feature information detection subunit includes: a unit embedding vector transformation module, used to convert the target image features into multi-dimensional unit embedding vectors; a species similarity suppression result determination module, used to perform species similarity suppression operation based on the unit embedding vectors to obtain the species similarity suppression result, wherein the species similarity suppression operation is used to enhance the distinguishability between the target object features and the features of closely related species through metric learning and interval classification strategies, so as to form a classification boundary in the embedding space; and a feature information detection module, used to detect feature information related to the target object in the image frame based on conditional features, species similarity suppression result and target image features.

[0146] Optionally, the feature information detection module includes: a contrastive learning loss function setting submodule, used to set contrastive learning loss functions for the embedded feature information in the target image bounding box and the embedded feature information in the region of interest, respectively, based on the target image features; and a feature information detection submodule, used to detect feature information related to the target object in the image frame based on conditional features, species similarity suppression results, target image features, and the contrastive learning loss functions set for the embedded feature information in the target image bounding box and the embedded feature information in the region of interest, respectively.

[0147] Optionally, the target image feature determination module includes: a third image feature determination submodule, used to perform weighted calculation on the first image feature and the second image feature based on the fusion weight to obtain the third image feature; and a target image feature determination submodule, used to perform a gating operation on the third image feature to obtain the target image feature, wherein the gating operation is used to suppress feature information in the third image feature that is lower than a preset frequency, and retain feature information in the third image feature that is not lower than a preset frequency.

[0148] Optionally, the species similarity suppression result determination module includes: an online maintenance submodule for maintaining the momentum prototype vectors of the target object and closely related species online, wherein the momentum prototype vector is a vector representing the features of the target object maintained through a momentum update mechanism, used to reflect the statistical characteristics of the target object features changing over time during metric learning; a cosine similarity determination submodule for calculating the cosine similarity between the unit embedding vector and the momentum prototype vector; and a classification boundary processing submodule for enhancing the discriminative power between the features of the target object and the features of closely related species based on cosine similarity and an interval classification strategy, so as to form a classification boundary in the embedding space, wherein the interval classification strategy is used to enhance the discriminative power of image feature representation by expanding the decision boundary between different categories in the embedding space.

[0149] Optionally, the contrastive learning loss function setting submodule includes: a first function determination submodule, used to set a region-level contrastive learning loss function corresponding to the embedded feature information in the region of interest based on the target image features and the embedded feature information in the region of interest, wherein the region-level contrastive learning loss function is used to expand the inter-class margin and compress the intra-class variance at the candidate region scale; and a second function determination submodule, used to set an image-level contrastive learning loss function corresponding to the embedded feature information in the target image bounding box based on the target image features and the embedded feature information in the target image bounding box, wherein the image-level contrastive learning loss function uses paired images within the same geofence and the same time window as positive samples, and images with different geofences / time windows or pure background images as negative samples.

[0150] Optionally, the target object detection device further includes: a target function determination unit, used to construct a target contrast learning loss function based on a region-level contrast learning loss function and an image-level contrast learning loss function through a multi-task learning mechanism; and a total loss function construction unit, used to construct a total loss function based on the target contrast learning loss function, the loss function set during the execution of the species similarity suppression operation, and the classification loss function and regression loss function set for the target object detection task.

[0151] Optionally, the second processing unit 903 includes: a parameter determination subunit, used to generate channel scale scaling parameters and bias parameters using conditional vectors; wherein the channel scale scaling parameters are learnable parameters used to dynamically adjust the response intensity of each channel in the feature map of the convolutional neural network, and the bias parameters are learnable parameters used to offset the response value of the feature map of the convolutional network channel by channel; and a modulation processing subunit, used to perform channel-by-channel linear modulation on the target convolutional layer of the convolutional neural network and the target network layer of the feature pyramid network through the channel scale scaling parameters and bias parameters to obtain conditional features.

[0152] Optionally, the third processing unit 904 includes: a candidate image box generation subunit, used to generate multiple candidate image boxes at multiple scales based on image frames using conditional features as input and a region proposal network; and a target image box determination subunit, used to filter target image boxes from multiple candidate image boxes according to the size and proportion of each candidate image box, wherein the size and proportion of the target image box match the size and proportion of the target object.

[0153] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, which stores a computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located performs the above-described target object detection method.

[0154] According to another aspect of the embodiments of this application, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by one or more processors, the one or more processors cause the one or more processors to perform the target object detection method described above.

[0155] According to another aspect of the embodiments of this application, a computer program product is also provided, including a computer program or instructions, which implement the above-described target object detection method when executed by a processor.

[0156] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0157] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0158] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0159] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0160] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0161] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0162] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for detecting a target object, characterized in that, include: Acquire positioning data and environmental data synchronized with image frames, wherein the image frames are image frames acquired in the living environment of the target object; The positioning data and the environmental data are mapped into a conditional vector using a multilayer perceptron. The conditional vector is used to perform channel-wise linear modulation on the target convolutional layer of the convolutional neural network and the target network layer of the pyramid network during the feature extraction process to obtain conditional features. The target convolutional layer is used to extract image features, and the target network layer is used to generate multi-scale feature maps. Based on the conditional features, multiple candidate image boxes are generated on the image frame, and then the target image box is obtained by filtering from the multiple candidate image boxes according to the size and proportion of each candidate image box. Detect feature information related to the target object in the image frame based on the target image bounding box; The method of detecting feature information related to the target object in the image frame based on the target image bounding box includes: drawing the target image bounding box to obtain a region of interest (ROI) within the target image bounding box; extracting image features from the ROI as a first image feature; proportionally expanding the first image feature based on the side length of the target image bounding box to obtain a second image feature; fusing the first image feature and the second image feature into a target image feature; and detecting feature information related to the target object in the image frame based on the target image feature. The method of fusing the first image feature and the second image feature into a target image feature includes: obtaining a global convergence vector between the first image feature and the second image feature, wherein the global convergence vector is used to compress the first image feature and the second image feature as multi-dimensional spatial features into a fixed-length vector; inputting the global convergence vector into a gated neural network to obtain fusion weights between the first image feature and the second image feature determined by the gated neural network based on contextual attention and frequency domain noise reduction mechanisms; and fusing the first image feature and the second image feature into a target image feature according to the fusion weights. Detecting feature information related to the target object in the image frame based on the target image features includes: converting the target image features into multi-dimensional unit embedding vectors; performing a species similarity suppression operation based on the unit embedding vectors to obtain a species similarity suppression result, wherein the species similarity suppression operation is used to enhance the distinguishability between the target object features and the features of closely related species through metric learning and interval classification strategies, so as to form a classification boundary in the embedding space; and detecting feature information related to the target object in the image frame based on the conditional features, the species similarity suppression result, and the target image features.

2. The target object detection method according to claim 1, characterized in that, Based on the conditionalized features, the species similarity suppression results, and the target image features, detect feature information related to the target object in the image frame, including: Based on the target image features, contrastive learning loss functions are set for the embedded feature information in the target image bounding box and the embedded feature information in the region of interest, respectively. Based on the conditional features, the species similarity suppression results, the target image features, and the contrastive learning loss functions set for the embedded feature information in the target image frame and the embedded feature information in the region of interest, feature information related to the target object in the image frame is detected.

3. The target object detection method according to claim 1, characterized in that, The first image feature and the second image feature are fused into a target image feature according to the fusion weight, including: The first image feature and the second image feature are weighted and calculated based on the fusion weight to obtain the third image feature; A gating operation is performed on the third image feature to obtain the target image feature, wherein the gating operation is used to suppress feature information in the third image feature that is below a preset frequency, and retain feature information in the third image feature that is not below the preset frequency.

4. The target object detection method according to claim 1, characterized in that, The species similarity suppression operation includes: The momentum prototype vectors of the target object and closely related species are maintained online. The momentum prototype vector is a vector representing the characteristics of the target object that is maintained through a momentum update mechanism. It is used to reflect the statistical characteristics of the target object's characteristics as they change over time during the metric learning process. Calculate the cosine similarity between the unit embedding vector and the momentum prototype vector; The discriminative power of the target object features and closely related species features is enhanced by the cosine similarity and the interval classification strategy to form a classification boundary in the embedding space. The interval classification strategy is used to enhance the discriminative power of image feature representations by expanding the decision boundary between different categories in the embedding space.

5. The target object detection method according to claim 2, characterized in that, Based on the target image features, contrastive learning loss functions are set for the embedded feature information in the target image bounding box and the embedded feature information in the region of interest, respectively, including: Based on the target image features and the embedded feature information in the region of interest, a region-level contrastive learning loss function corresponding to the embedded feature information in the region of interest is set, wherein the region-level contrastive learning loss function is used to expand the inter-class interval and compress the intra-class variance at the candidate region scale; Based on the target image features and the embedded feature information in the target image bounding box, an image-level contrastive learning loss function corresponding to the embedded feature information in the target image bounding box is set. The image-level contrastive learning loss function uses paired images in the same geofence and the same time window as positive samples, and images in different geofences / time windows or pure backgrounds as negative samples.

6. The target object detection method according to claim 5, characterized in that, The target object detection method further includes: The target contrast learning loss function is constructed by using a multi-task learning mechanism based on the region-level contrast learning loss function and the image-level contrast learning loss function. The total loss function is constructed based on the target contrast learning loss function, the loss function set during the execution of the species similarity suppression operation, and the classification loss function and regression loss function set for the target object detection task.

7. The target object detection method according to claim 5, characterized in that, The positive samples corresponding to the region-level contrastive learning loss function are multiple first-class regions of interest from the same real target, and the negative sample pairs corresponding to the region-level contrastive learning loss function are multiple regions of interest from different species or from second-class regions of interest. The cross-union ratio of the first-class regions of interest is greater than a first threshold, and the cross-union ratio of the second-class regions of interest is less than a second threshold, wherein the first threshold is greater than the second threshold.

8. The target object detection method according to claim 1, characterized in that, Using the conditional vector, channel-wise linear modulation is performed on the target convolutional layer of the convolutional neural network and the target network layer of the pyramid network during the feature extraction process to obtain conditional features, including: The conditional vector is used to generate channel scaling parameters and bias parameters; wherein, the channel scaling parameters are learned parameters used to dynamically adjust the response intensity of each channel in the feature map of the convolutional neural network, and the bias parameters are learned parameters used to offset the response value of the convolutional network feature map channel by channel. The target convolutional layer of the convolutional neural network and the target network layer of the feature pyramid network are subjected to channel-wise linear modulation using the channel scale scaling parameter and the bias parameter to obtain the conditional features.

9. The target object detection method according to claim 1, characterized in that, Based on the conditional features, multiple candidate image boxes are generated on the image frame. Then, the target image box is obtained by filtering from the multiple candidate image boxes according to the size and proportion of each candidate image box, including: Using conditional features as input, a region proposal network generates multiple candidate image boxes at multiple scales based on the image frame. A target image frame is obtained by filtering from the plurality of candidate image frames based on the size and proportion of each candidate image frame, wherein the size and proportion of the target image frame are matched with the body size and proportion of the target object.

10. A target object detection device, used to implement the target object detection method according to any one of claims 1 to 9, characterized in that, include: The acquisition unit is used to acquire positioning data and environmental data synchronized with the image frame, wherein the image frame is an image frame acquired in the living environment of the target object; The first processing unit is used to map the positioning data and the environmental data into a conditional vector using a multilayer perceptron. The second processing unit is used to perform channel-wise linear modulation on the target convolutional layer of the convolutional neural network and the target network layer of the pyramid network during the feature extraction process using the conditional vector to obtain conditional features, wherein the target convolutional layer is used to extract image features and the target network layer is used to generate multi-scale feature maps. The third processing unit is used to generate multiple candidate image boxes on the image frame based on the conditional features, and then filter the target image box from the multiple candidate image boxes according to the size and proportion of each candidate image box. The detection unit is used to detect feature information related to the target object in the image frame based on the target image frame.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein when the computer program is executed, the device on which the computer-readable storage medium is located performs the target object detection method according to any one of claims 1 to 9.

12. An electronic device, characterized in that, It includes one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to perform the target object detection method according to any one of claims 1 to 9.

13. A computer program product, characterized in that, It includes a computer program or instructions that, when executed by a processor, implement the target object detection method according to any one of claims 1 to 9.

Images