Tracking camera low-power-consumption technical method and device based on edge AI and camera

Abstract

The invention discloses a tracking camera low-power-consumption technical method and device based on edge AI and a camera, and relates to the field of image recognition, and the method comprises the steps: deploying a target recognition model for recognizing a target object at a camera end; starting image data acquisition, and inputting the acquired image data into the target recognition model to obtain a candidate area; performing feature matching according to the candidate area and the target object, judging whether the target object exists in the image data, and obtaining a judgment result; and when the judgment result is that the target object does not exist in the image data, controlling the camera end to terminate image data acquisition and enter a low-power-consumption state. According to the method, the lightweight AI model is deployed at the camera end, real-time intelligent power consumption control based on image semantic understanding is achieved, invalid scenes can be actively recognized, instant dormancy can be achieved, and energy waste caused by recording invalid content in a traditional scheme is fundamentally eliminated.

Description

Technical Field

[0001] This invention relates to the field of image recognition, and more particularly to a low-power technology method, device, and camera for tracking cameras based on edge AI. Background Technology

[0002] With the increasing popularity of outdoor adventure, ecological observation, and wilderness security activities, tracking cameras, as an important information collection device, are being used more and more widely. Since outdoor environments typically lack a stable mains power supply, the vast majority of these devices rely on built-in batteries to operate. Therefore, the power consumption level of the device directly determines its battery life per deployment, becoming a core bottleneck affecting user experience and practicality.

[0003] To extend battery life, the industry has widely adopted various low-power technologies, but all have significant drawbacks: The first is the timed shooting solution; this solution cannot sense scene content, resulting in periodic shooting even during periods without a target, generating a large amount of invalid data and power consumption. The second is a motion detection solution based on passive infrared sensors; this solution can only detect heat source movement, cannot identify the type of moving object, and is easily affected by environmental thermal disturbances, small animals, and other non-target interference, resulting in false triggers and invalid recordings. The fundamental flaw of existing technologies lies in the fact that their triggering mechanisms are severely disconnected from the core user need of "only recording valuable specific targets." Whether it is the indiscriminate shooting of timers or the non-identification triggering of passive infrared sensors, their essence is non-intelligent decision-making based on simple physical signals, unable to judge the value of the content at the source of acquisition. This leads to a large amount of the device's precious power being consumed in recording and processing empty scenes or non-target objects, forming a bottleneck that is difficult to overcome in terms of battery life. Summary of the Invention

[0004] The main objective of this invention is to provide a low-power technology method, device, and camera for tracking cameras based on edge AI, aiming to solve the technical problem that a large amount of power is consumed in recording and processing aerial scenes or non-target objects.

[0005] To achieve the above objectives, this invention provides a low-power tracking camera technology method based on edge AI, the method comprising: S100 deploys a target recognition model at the camera end to identify target objects; S200: Start image data acquisition and input the acquired image data into the target recognition model to obtain candidate regions; S300: Based on feature matching between candidate regions and target objects, determine whether a target object exists in the image data and obtain the determination result; S400: When the judgment result is that there is no target object in the image data, the camera is controlled to terminate image data acquisition and enter a low power state.

[0006] This invention achieves real-time intelligent power consumption control based on image semantic understanding by deploying a lightweight AI model on the camera end. It can actively identify invalid scenes and instantly put the camera into sleep mode, fundamentally eliminating the energy waste caused by recording invalid content in traditional solutions.

[0007] Furthermore, step S100 also includes model training and deployment steps: The initial neural network model is trained using a large number of sample images containing the target object to generate a target recognition model; The trained target recognition model is quantized and compressed to fit the computing resources and storage space of the camera. The quantized and compressed model is deployed to the camera.

[0008] Furthermore, initiating image data acquisition includes preprocessing the raw image of the target object, which includes: Image data of different formats are uniformly converted into the standard format preset by the model and pixel values ​​are normalized; Adaptive median filtering is used to remove noise from the image data; lens distortion correction is performed on the target image data based on the intrinsic parameter matrix at the camera end. Illumination equalization of image data is achieved through histogram equalization or the Retinex algorithm; color deviation of image data under different light sources is corrected based on white balance technology.

[0009] Furthermore, inputting the acquired image data into the target recognition model to obtain candidate regions includes: S210, The preprocessed image is scanned using a target detection algorithm to locate one or more candidate regions containing the target object; S220, Perform feature extraction on each candidate region to obtain the feature vector of the candidate region; S230, calculate the confidence level of the feature vector of each candidate region, and select the candidate regions with confidence levels higher than the preset threshold as candidate regions.

[0010] Furthermore, feature extraction methods include: S221, extracts multi-level visual features of candidate regions from image data through a deep convolutional neural network of the target recognition model; S222 uses an attention mechanism to enhance the weights of feature channels or spatial regions related to the target object in the candidate region and suppress background features to obtain the enhanced high-dimensional features of the candidate region. S223 uses a fully connected layer of the target recognition model to reduce and compress the enhanced high-dimensional features to obtain the feature vector of the candidate region.

[0011] Furthermore, the method for determining whether a target object exists in image data includes: setting a preset similarity threshold; calculating the similarity between the feature vectors of candidate regions and the target object standard feature library built into the model; if the similarity between the feature vector of any candidate region and any feature in the target object standard feature library exceeds the similarity threshold, then the image is determined to contain a target object; if the similarity between the feature vectors of all candidate regions and all features in the target object standard feature library is lower than the similarity threshold, then the image is determined to contain no target object.

[0012] Furthermore, the target object standard feature library is constructed and updated in the following way: During the model training phase, feature vectors obtained by passing all positive samples in the training set through the feature extraction network are extracted to form the initial standard feature library; During the actual operation at the camera end, feature vectors of new images that are confirmed as target objects and have a confidence level higher than a preset update threshold are added to the local standard feature library to achieve online incremental learning of features.

[0013] Furthermore, when the judgment result indicates that a target object exists in the image data, the method also includes: starting a continuous recording mode and inputting the image data during the continuous recording process into the target recognition model in real time or near real time; If no target object is detected within a preset number of frames or a preset duration during continuous recording, the camera will stop recording and enter a low-power state.

[0014] The beneficial effects of this invention are as follows: by deploying a lightweight AI model on the camera end, real-time intelligent power consumption control based on image semantic understanding is realized. It can actively identify invalid scenes and instantly put the camera into sleep mode, fundamentally eliminating the energy waste caused by recording invalid content in traditional solutions.

[0015] This invention also provides a low-power tracking camera technology device based on edge AI. The control device includes: a memory, a processor, and a computer program stored in the memory and executed by the processor. When executed by the processor, the computer program implements a low-power tracking camera technology method based on edge AI. The present invention also proposes a camera that includes the aforementioned low-power tracking camera device based on edge AI. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating an embodiment of the low-power tracking camera technology method based on edge AI of the present invention. Figure 2 This is a schematic diagram illustrating the steps for obtaining candidate regions in an embodiment of the low-power tracking camera technology method based on edge AI in this invention. Figure 3 This is a schematic diagram of the feature extraction and matching judgment process provided in an embodiment of the low-power tracking camera technology method based on edge AI in this invention. Figure 4 This is a schematic diagram of the module structure of the low-power tracking camera technology system based on edge AI according to an embodiment of the present invention; The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of the present invention and are not intended to limit the present invention.

[0020] To better understand the technical solution of the present invention, a detailed description will be provided below in conjunction with the accompanying drawings and specific embodiments.

[0021] The main solution of this application is as follows: deploy a target recognition model for identifying target objects on the camera end; start image data acquisition and input the acquired image data into the target recognition model to obtain candidate regions; perform feature matching between the candidate regions and the target objects to determine whether the target objects exist in the image data and obtain the judgment result; when the judgment result is that the target objects do not exist in the image data, control the camera end to terminate image data acquisition and enter a low power state.

[0022] Currently, outdoor environments typically lack readily available power sources, and most camera devices rely on batteries for power. Therefore, the power consumption of a camera device directly determines its battery life, becoming a key factor affecting user experience. Given limited battery capacity, effectively reducing power consumption and extending battery life has become a core technical challenge for the industry. By deploying an AI model on the local camera, real-time content recognition can be performed at the initial stage of image acquisition. Once it is determined that no target object exists in the scene, acquisition is immediately terminated and a low-power state is entered, fundamentally avoiding the energy waste caused by recording invalid video and greatly improving the device's battery life.

[0023] Based on this, this application proposes an embodiment of a low-power tracking camera technology method based on edge AI. Please refer to [link / reference]. Figure 1 The low-power technology method for tracking cameras based on edge AI includes steps S100~S400. Furthermore, in step S100, the method for deploying the target recognition model for identifying the target object at the camera end is as follows: In this embodiment, a deep convolutional neural network is trained in the cloud using a large-scale labeled dataset, such as tens of thousands of animal images under different lighting, angles, and poses. After training, redundant convolutional kernels are removed by pruning, and the model weights are converted from FP32 to INT8 precision by quantization. Then, knowledge distillation is used to guide the training of a smaller model using the larger model, ultimately generating a lightweight model with a size reduction of approximately 75% and a computational reduction of approximately 80%. This model is downloaded to the camera's Flash memory via wired or wireless means and loaded into the processor's memory for execution upon startup.

[0024] In this embodiment, model lightweighting is a key prerequisite for its real-time operation on resource-constrained edge devices. Pruning and quantization directly reduce the number of model parameters and computational accuracy, thereby reducing memory usage and CPU / GPU computational burden. Knowledge distillation, on the other hand, compresses the model while preserving as much of the original model's recognition capabilities as possible. Through the comprehensive application of these technologies, complex AI models can run in real-time with low power consumption on camera-embedded chips, ensuring the feasibility of "real-time analysis" and making the power consumption of AI inference itself far lower than the power savings achieved through intelligent control, resulting in positive benefits.

[0025] Furthermore, in step S200, the method for initiating image data acquisition and inputting the acquired image data into the target recognition model to obtain candidate regions is as follows: In this embodiment, the camera is woken up every 2 seconds by a timer, and the image sensor acquires a frame of RAW format image. Preprocessing is performed sequentially: RAW is converted to RGB through the ISP pipeline, and the pixel values ​​are normalized from [0,255] to [0,1]; noise reduction is performed using an adaptive median filter with a 5x5 window, and lens distortion correction is performed using a pre-calibrated camera intrinsic parameter matrix for focal length, principal point, and distortion coefficients; local histogram equalization is performed using the CLAHE algorithm to improve contrast, and a white balance algorithm based on the white balance algorithm is used to correct color cast. The CLAHE algorithm is a contrast-limited adaptive histogram equalization algorithm used to improve local image contrast.

[0026] In another feasible implementation, waking up the camera can be initially triggered by a passive infrared sensor.

[0027] In another feasible implementation, improving image contrast and illumination balance can be accomplished using the Retinex algorithm.

[0028] In this embodiment, outdoor environments are characterized by variable lighting and strong noise interference, directly impacting recognition accuracy due to the quality of the original image. This step employs a series of standardization processes to transform the original image data collected under different conditions into image data with uniform quality and consistent features. For example, distortion correction ensures that the shape of the main target in the image data remains unchanged, and white balance eliminates the influence of color temperature caused by varying lighting conditions, making the target's color features more comparable to those in the training library. This significantly improves the robustness and accuracy of subsequent feature extraction and matching, while higher recognition accuracy reduces the probability of false triggering and false shutdown, thus fundamentally ensuring the effectiveness of energy-saving measures.

[0029] like Figure 2 As shown, the steps for inputting the acquired image data into the target recognition model to obtain candidate regions include: S210: The preprocessed image is scanned using a target detection algorithm to locate one or more candidate regions containing the target object.

[0030] In this embodiment, the preprocessed image is input to the front end of the target recognition model. This front end is based on a lightweight single-stage detection algorithm such as YOLOv5-nano. It performs dense sliding window detection on the feature map and outputs multiple bounding boxes. The regions within the bounding boxes are the "candidate regions", and each box is accompanied by an initial class confidence score, such as "animal: 0.6".

[0031] S220: Extract features from each candidate region to obtain the feature vector of the candidate region.

[0032] After obtaining the candidate regions, more refined feature analysis is performed on them. For example... Figure 3As shown, this step is the core of improving recognition accuracy, and specifically includes: S221 extracts multi-level visual features of candidate regions from image data through a deep convolutional neural network of the target recognition model.

[0033] Specifically, a fixed-size feature map patch is extracted from the original feature map for each candidate region. This feature map patch is then input into a feature extraction sub-network, which is a small deep convolutional neural network, such as a MobileNet block. The deep convolutional neural network abstracts layer by layer from the input, extracting multi-level visual features, from low-level edges and textures to high-level semantic structures, forming a rich feature tensor.

[0034] S222, through an attention mechanism, the feature channels or spatial regions related to the target object in the candidate region are weighted and enhanced, while background features are suppressed, to obtain the enhanced high-dimensional features of the candidate region. To improve the representational and discriminative power of the features, this embodiment introduces an attention mechanism during feature extraction. Specifically, in this embodiment, the attention mechanism introduces a channel attention module. The attention module obtains global information for each channel through global average pooling, learns the importance weights of each channel through two fully connected layers, and finally performs channel weighting on the original features. For example, for the target "deer", the attention mechanism automatically assigns higher weights to feature channels related to "antler texture" and "fur color", while suppressing channels related to the "leaf" background.

[0035] Specifically, in this embodiment, the attention mechanism calculates the importance of different spatial locations in the feature map and generates an attention heatmap, thereby making the network pay more attention to the area where the target object is most likely to appear within the candidate region and reducing the interference of the background region.

[0036] Through the synergistic effect of one or more of the above attention mechanisms, the network can focus on key information, thereby obtaining enhanced and more discriminative feature tensors.

[0037] S223 uses a fully connected layer of the target recognition model to reduce and compress the enhanced high-dimensional features to obtain the feature vector of the candidate region.

[0038] In this embodiment, the attention-enhanced feature tensor is high-dimensional and contains information redundancy. The feature tensor is flattened and input into a fully connected layer. The fully connected layer maps the high-dimensional features to a lower-dimensional vector space, thereby reducing computational and storage overhead, and integrating information from all spatial locations and channels to form the feature vector of the candidate region.

[0039] In this embodiment, the feature vector of the candidate region is an abstract mathematical representation of the candidate region, which contains its most essential and discriminative visual information and is used for matching the target object.

[0040] S230, calculate the confidence level of the feature vector of each candidate region, and select the candidate regions with confidence levels higher than the preset threshold as candidate regions.

[0041] In this embodiment, the target recognition model evaluates the confidence level of each feature vector obtained in S220. The confidence level is a value between 0 and 1, which quantifies the probability that the candidate region contains the target object. Specifically, the target recognition model uses a lightweight classifier to output a confidence score of the feature vector belonging to each category based on the input feature vector.

[0042] Furthermore, in this embodiment, the system performs non-maximum suppression: all candidate regions are sorted in descending order according to their confidence level; duplicate regions whose overlap with the highest confidence region exceeds a set threshold (0.5) are removed; finally, regions with confidence levels higher than a preset threshold (0.7) are retained from the remaining regions. The high-quality regions retained after filtering are recorded as candidate regions. This step can effectively filter out low-quality false positives and duplicate detection boxes.

[0043] Furthermore, in step S300, feature matching is performed between the candidate region and the target object to determine whether the target object exists in the image data, and the determination result is obtained, including: Specifically, during system initialization or model deployment, a similarity threshold (0.85) needs to be preset. This similarity threshold is used to determine whether feature matching is successful. Furthermore, the similarity threshold setting needs to strike a balance between accuracy and recall: a threshold that is too high may lead to missed detections, where even if a target exists, it is judged as not being a target due to insufficient similarity; a threshold that is too low may increase false detections, misclassifying similar non-target objects as targets.

[0044] For example, the similarity threshold can be set empirically or optimized and calibrated using a validation set, depending on the specific type of the target object, the complexity of the environment, and the accuracy requirements of the application scenario.

[0045] Specifically, the target object standard feature library is the knowledge foundation for the accurate identification achieved by this invention, and it consists of a set of feature vectors obtained by processing a large number of target object samples through a feature extraction network. This set contains diverse feature representations of target objects under different angles, lighting, scales, and partial occlusion, and serves as a robust reference standard.

[0046] Furthermore, for each candidate region's feature vector, the system calculates its similarity to all feature vectors in the standard feature library or the most relevant feature vectors retrieved by index.

[0047] Based on all calculated similarity results, the judgment logic is as follows: If, among all calculated similarities, at least one candidate region's feature vector has a similarity value exceeding a preset similarity threshold with any feature vector in the standard feature library, then the system determines that a target object exists in the current image data. This indicates that the camera has successfully captured an object highly similar to a known target.

[0048] If the similarity values ​​between the feature vectors of all candidate regions and all feature vectors in the standard feature library are lower than a preset similarity threshold, the system determines that there is no target object in the current image data. This indicates that no object in the current image meets the feature criteria of the target object.

[0049] In this embodiment, the similarity calculation uses a cosine similarity algorithm. The system pre-stores a standard feature library for target objects, which consists of a large number of positive sample feature vectors extracted during the training phase. The 256-dimensional feature vector of each candidate region selected in S230 is compared with the vectors in the feature library using cosine similarity calculation. A similarity threshold θ is set, where θ = 0.85. If the similarity between any feature of a candidate region and any feature in the library is greater than θ, the candidate region is determined to contain a target object. If all similarities are less than θ, the candidate region is determined to contain no target object. Using a local feature library for comparison enables completely offline recognition, eliminating the dependence on network connectivity. This makes it suitable for signal-free outdoor environments and offers the advantages of zero network latency and zero data consumption.

[0050] Furthermore, in step S400, when the determination result is that there is no target object in the image data, the camera is controlled to terminate image data acquisition and enter a low-power state.

[0051] In this embodiment, the control logic based on the judgment result of S300 is as follows: If it is determined that there is no target object in the current frame, the control logic will immediately send an instruction to the camera system to terminate the current photo or video recording process, immediately stop the image sensor exposure, turn off the encoder, and put the main control CPU into a deep sleep mode, thereby achieving the fastest energy-saving response, which is much faster than the traditional fixed-duration recording end and then sleep mode.

[0052] If a target is detected, the camera enters continuous recording mode; simultaneously, the S200-S300 recognition process runs continuously in the background at once per second. If no target is detected within 30 consecutive frames during subsequent recording, the system determines that the target has left, automatically stops recording, and enters a low-power state or sleep state.

[0053] This judgment mechanism enables dynamic and adaptive power management strategies. Immediate termination avoids the target briefly spinning after a bird flies by; delayed termination avoids frequent start-stop cycles caused by the target object briefly obscuring or going out of frame, greatly improving the user experience. This intelligent control mechanism based on continuous content analysis can accurately match the target's activity patterns in real-world scenarios, strictly synchronizing the device's operating status with the generation of effective information, and achieving ultimate power optimization.

[0054] like Figure 4 As shown, Figure 4 This is a schematic diagram of the module structure of the low-power tracking camera technology device based on edge AI according to an embodiment of the present invention.

[0055] Reference Figure 4 The present invention also proposes a low-power tracking camera device based on edge AI. The device includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of a low-power tracking camera method based on edge AI. The device includes: Deployment module 10 is used to deploy a target recognition model for identifying target objects at the camera end; The acquisition module 20 is used to initiate image data acquisition and input the acquired image data into the target recognition model to obtain candidate regions; The matching module 30 is used to perform feature matching between candidate regions and target objects, determine whether a target object exists in the image data, and obtain the determination result. The judgment module 40 is used to control the camera to terminate image data acquisition and enter a low-power state when the judgment result is that the target object does not exist in the image data.

[0056] Those skilled in the art will understand that the example is merely an example of a low-power tracking camera device based on edge AI and does not constitute a limitation on a low-power tracking camera device based on edge AI. It may include more or fewer components than the example, or combine certain components, or different components. For example, a low-power tracking camera device based on edge AI may also include input / output devices, network access devices, buses, etc.

[0057] By implementing a low-power tracking camera technology device based on edge AI, and by deploying a lightweight AI model at the camera end, real-time intelligent power consumption control based on image semantic understanding is achieved. It can actively identify invalid scenes and instantly put the camera into sleep mode, fundamentally eliminating the energy waste caused by recording invalid content in traditional solutions.

[0058] In this embodiment, the processor executes a low-power tracking camera technology method based on edge AI stored in the memory. By deploying a lightweight AI model at the camera end, real-time intelligent power consumption control based on image semantic understanding is achieved. It can actively identify invalid scenes and instantly put the camera into sleep mode, fundamentally eliminating the energy waste caused by recording invalid content in traditional solutions.

[0059] The present invention also proposes a low-power tracking camera technology device based on edge AI, the device comprising: a memory, a processor, and a computer program stored on the memory and executable on the processor, the computer program being configured to implement the steps of the low-power tracking camera technology method based on edge AI as described in any of the first aspects above.

[0060] The above are only some embodiments of the present invention and do not limit the patent scope of the present invention. All equivalent structural transformations made under the technical concept of the present invention using the contents of the present invention specification and drawings, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A low-power tracking camera technique based on edge AI, the method comprising: S100 deploys a target recognition model at the camera end to identify target objects; S200, initiate image data acquisition and input the acquired image data into the target recognition model to obtain candidate regions; S300: Based on feature matching between candidate regions and target objects, determine whether a target object exists in the image data and obtain the determination result; S400, when the judgment result is that there is no target object in the image data, control the camera to terminate image data acquisition and enter a low power consumption state.

2. The method according to claim 1, characterized in that, The target recognition model is a lightweight model that has undergone quantization and compression.

3. The method according to claim 1, characterized in that, The initiation of image data acquisition includes: preprocessing the original image of the target object, wherein the preprocessing includes at least one of the following: Image data of different formats are uniformly converted into the model's preset standard format and pixel values ​​are normalized; adaptive median filtering is used to remove noise from the image data; lens distortion correction is performed on the target image data based on the camera's intrinsic parameter matrix; illumination equalization is performed on the image data through histogram equalization or the Retinex algorithm; and color deviation of the image data under different light sources is corrected based on white balance technology.

4. The method according to claim 1, characterized in that, The step of inputting the collected image data into the target recognition model to obtain candidate regions includes: S210, The preprocessed image is scanned using a target detection algorithm to locate one or more candidate regions containing the target object; S220, Perform feature extraction on each candidate region to obtain the feature vector of the candidate region; S230, calculate the confidence level of the feature vector of each candidate region, and select the candidate regions with confidence levels higher than the preset threshold as candidate regions.

5. The method according to claim 4, characterized in that, The method for extracting features from each of the candidate regions to obtain the feature vector of the candidate region includes: S221, extracts multi-level visual features of candidate regions from image data through a deep convolutional neural network of the target recognition model; S222 uses an attention mechanism to enhance the weights of feature channels or spatial regions related to the target object in the candidate region and suppress background features to obtain the enhanced high-dimensional features of the candidate region. S223 uses a fully connected layer of the target recognition model to reduce and compress the enhanced high-dimensional features to obtain the feature vector of the candidate region.

6. The method according to claim 4 or 5, characterized in that, The method for determining whether a target object exists in image data by performing feature matching between candidate regions and target objects includes: Set a preset similarity threshold; The similarity between the feature vectors of the candidate regions and the standard feature library of the target object built into the model is calculated to obtain the similarity between the feature vector of each candidate region and each feature of the standard feature library of the target object. If the similarity between the feature vector of any candidate region and any feature in the standard feature library of the target object exceeds the similarity threshold, then the image is determined to contain a target object. If the similarity between the feature vectors of all candidate regions and all features in the standard feature library of the target object is lower than the similarity threshold, then it is determined that there is no target object in the image.

7. The method according to claim 6, characterized in that, The target object standard feature library is constructed and updated in the following ways: During the model training phase, feature vectors are extracted from all positive samples in the training set after passing through the feature extraction network to form the initial standard feature library. During actual operation on the camera, the feature vectors of new images that are confirmed as target objects and have a confidence level higher than a preset update threshold are added to the local standard feature library.

8. The method according to claim 1, characterized in that, When the determination result indicates that a target object exists in the image data, the method further includes: Start the continuous recording mode and input the image data during the continuous recording process into the target recognition model; If no target object is detected within a preset number of frames or a preset duration during continuous recording, the camera will be controlled to terminate recording and enter the low-power state.

9. A low-power technology device for tracking cameras based on edge AI, characterized in that, The control device includes: a memory, a processor, and a computer program stored in the memory and executed by the processor, wherein when the computer program is executed by the processor, it implements the low-power technology method for tracking cameras based on edge AI as described in any one of claims 1-8.

10. A camera, characterized in that, The camera includes the low-power tracking camera device based on edge AI as described in claim 9.

Images