Small target tracking method and system under low illumination condition, storage medium and equipment

Abstract

The embodiment of the invention discloses a small target tracking method under a low illumination condition, and the method comprises the steps: obtaining an image frame by frame, carrying out the recognition of a target through a target detector according to the image, and determining the target position and target feature information of a current frame; processing the target position and the target feature information of the current frame based on an En-GhostNet network to obtain a target appearance feature; predicting a next frame target position corresponding to the target appearance feature by using Kalman filtering; and outputting the tracking information of the target according to frames. According to the method, the current frame target is identified through target detection, the appearance characteristics of the target are extracted through the En-GhostNet network, finally, the appearance characteristics of the target are processed by using Kalman filtering to obtain the motion trail of the target, and finally, accurate tracking of the small target under the low-light condition is realized.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a small target tracking method, system, storage medium and device under low-light conditions. Background Art

[0002] Image quality issues in low-light environments pose numerous challenges to the field of computer vision, particularly in target tracking. Due to insufficient illumination, images often suffer from loss of detail in dark areas, increased noise, and color distortion, making target recognition and tracking difficult. Traditional target tracking methods, such as those based on template matching and optical flow, often fail to effectively capture the target's motion trajectory in low-light conditions, and are prone to tracking loss or mistracking.

[0003] In order to solve the problem of target tracking under low-light conditions, researchers have proposed various methods, mainly including two categories: image enhancement and feature extraction. Image enhancement methods, such as histogram equalization and contrast stretching, aim to improve the brightness and contrast of the image and enhance the visibility of the target. However, these methods may amplify background noise and reduce tracking accuracy. Feature extraction methods, such as HOG, SIFT, and SURF, track the target by extracting the texture and shape information of the target. However, these traditional feature extraction methods often lack sufficient robustness when dealing with challenges such as illumination changes, target appearance changes, occlusion, rapid motion, and complex background interference, making it difficult to ensure tracking accuracy and stability. Summary of the Invention

[0004] Based on this, it is necessary to propose a small target tracking method under low light conditions to address the above problems.

[0005] A small target tracking method under low light conditions, the method comprising the following steps:

[0006] Acquire images frame by frame, identify targets through target detectors, and determine the target position and target feature information of the current frame;

[0007] Processing the target position and target feature information of the current frame based on the En-GhostNet network to obtain target appearance features;

[0008] Using Kalman filtering to predict the target position in the next frame corresponding to the target appearance feature;

[0009] Output the tracking information of the target frame by frame.

[0010] In the above solution, the process of identifying the target in the image through a target detector and determining the target position and target feature information of the current frame specifically includes:

[0011] The target detector includes a perception feature recombination module and a small target detection layer;

[0012] Determine the current frame target position by the perception feature reconstruction module;

[0013] Target feature information is determined through the small target detection layer.

[0014] In the above solution, determining the current frame target position by the perception feature recombination module specifically includes:

[0015] Obtain the feature map corresponding to the target, and perform channel compression on the feature map through 1×1 convolution to generate a compressed feature map;

[0016] Determine an upsampling kernel of a specific size according to the compressed feature map;

[0017] Normalizing the upsampling kernel of the specific size to obtain a normalized upsampling kernel;

[0018] Perform element-by-element dot multiplication on the normalized up-sampling kernel and the feature map corresponding to the target to determine the target position of the current frame.

[0019] In the above solution, determining target feature information through the small target detection layer specifically includes:

[0020] Extracting four feature maps of the target at different resolutions through the small target detection layer, wherein the feature maps at different resolutions represent semantic information at different levels;

[0021] The four feature maps with different resolutions are fused through the feature pyramid network and the path aggregation network to obtain multi-scale information;

[0022] Convolution kernels of different sizes and types are used to perform convolution operations on multi-scale information to obtain target feature information.

[0023] In the above solution, the En-GhostNet network is used to process the target position and target feature information of the current frame to obtain the target appearance features, specifically including:

[0024] The En-GhostNet network includes a Ghost module and a Ghost bottleneck structure;

[0025] The Ghost module generates a corresponding feature map based on the current frame target position and target feature information;

[0026] Constructing a Ghost bottleneck structure, wherein the Ghost bottleneck structure includes a first Ghost layer and a second Ghost layer;

[0027] The first Ghost layer expands the number of channels through 1×1 convolution to obtain multi-dimensional detail information of the corresponding feature map;

[0028] The second Ghost layer compresses the number of channels through 1×1 convolution and converts the multi-dimensional detail information into target appearance features.

[0029] In the above solution, the use of Kalman filtering to predict the target position of the next frame corresponding to the target appearance feature specifically includes:

[0030] Determine the target state x according to the target appearance feature k , the target state x k Including the target position x k and speed v k :

[0031]

[0032] Predict the next frame state of the target according to the following formula:

[0033]

[0034] in, is the target state predicted by the current frame, F is the state transfer matrix, which indicates the change of the target state from the previous frame to the current frame, x k-1 is the target state of the previous frame;

[0035] The Euclidean distance between the target position and the predicted position is calculated according to the following formula:

[0036]

[0037] Among them, (x pred ,y pred ) is the predicted target position, (x det ,y det ) is the detected target position;

[0038] If the Euclidean distance between the target position and the predicted position meets the threshold range, the target position is updated using the current predicted position to determine the target position for the next frame:

[0039] x k =x pred +α·(x det -x pred )

[0040] Among them, x pred is the predicted position, x det is the current detection position, and α is the weight factor.

[0041] In the above solution, the outputting of the tracking information of the target frame by frame specifically includes:

[0042] The tracking information of the target includes the target's bounding box information, confidence value and corresponding tracking ID;

[0043] Determine the corresponding bounding box information according to the target position of the target in the current frame;

[0044] Assign a unique tracking ID to each target based on its location;

[0045] A confidence value is determined based on the Euclidean distance between the target position and the predicted position.

[0046] The present application also proposes a small target tracking system under low light conditions, the system comprising: a target detection unit, a feature extraction unit, a prediction unit and an output unit;

[0047] The target detection unit is used to acquire images frame by frame, identify the target in the image through the target detector, and determine the target position and target feature information of the current frame;

[0048] A feature extraction unit processes the target position and target feature information of the current frame based on the En-GhostNet network to obtain target appearance features;

[0049] A prediction unit, using Kalman filtering to predict the target position in the next frame corresponding to the target appearance feature;

[0050] The output unit outputs the tracking information of the target frame by frame.

[0051] The present application also provides a readable storage medium storing a computer program. When the computer program is executed by a processor, the processor performs the following steps:

[0052] Acquire images frame by frame, identify targets through target detectors, and determine the target position and target feature information of the current frame;

[0053] Processing the target position and target feature information of the current frame based on the En-GhostNet network to obtain target appearance features;

[0054] Using Kalman filtering to predict the target position in the next frame corresponding to the target appearance feature;

[0055] Output the tracking information of the target frame by frame.

[0056] The present application also proposes a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the following steps:

[0057] Acquire images frame by frame, identify targets through target detectors, and determine the target position and target feature information of the current frame;

[0058] Processing the target position and target feature information of the current frame based on the En-GhostNet network to obtain target appearance features;

[0059] Using Kalman filtering to predict the target position in the next frame corresponding to the target appearance feature;

[0060] Output the tracking information of the target frame by frame.

[0061] The embodiment of the present invention has the following beneficial effects: first, images are acquired frame by frame, the images are used to identify the target through a target detector, and the target position and target feature information of the current frame are determined; the target position and target feature information of the current frame are processed based on the En-GhostNet network to obtain the target appearance features; the target position of the next frame corresponding to the target appearance features is predicted using a Kalman filter; and the tracking information of the target is output frame by frame. The present invention identifies the target in the current frame through target detection, extracts the target appearance features using En-GhostNet, and predicts the target motion trajectory using a Kalman filter, ultimately achieving accurate tracking of small targets under low-light conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0063] in:

[0064] Figure 1 2. A flowchart of a method for tracking a small target under low light conditions according to an embodiment;

[0065] Figure 2 Schematic diagram of a content-aware feature reconstruction (CARAFE) module in one embodiment;

[0066] Figure 3 Schematic diagram of the network structure corresponding to the small target detection layer in one embodiment;

[0067] Figure 4 A schematic diagram of the Ghost module structure in one embodiment;

[0068] Figure 5 A schematic diagram of the Ghost bottleneck structure implementation. DETAILED DESCRIPTION

[0069] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0070] In the following description, numerous specific details are given to provide a more thorough understanding of the present invention; however, it will be apparent to those skilled in the art that the present invention may be practiced without one or more of these details; in other examples, some technical features known in the art are not described to avoid confusion with the present invention, and it should be understood that the present invention may be practiced in different forms and should not be construed as limited to the embodiments set forth herein; on the contrary, these embodiments are provided to make the disclosure thorough and complete and to fully convey the scope of the present invention to those skilled in the art.

[0071] The terminology used herein is intended only to describe specific embodiments and is not intended to limit the present invention. When used herein, the singular forms "a," "an," and "the" are intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the terms "comprising" and / or "including," when used in this specification, identify the presence of the recited features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups. When used herein, the term "and / or" includes any and all combinations of the relevant listed items.

[0072] In order to thoroughly understand the present invention, a detailed structure will be presented in the following description to illustrate the technical solution proposed by the present invention; optional embodiments of the present invention are described in detail below, but in addition to these detailed descriptions, the present invention may also have other implementation methods.

[0073] like Figure 1 As shown, in one embodiment, a small target tracking method under low light conditions is provided. The small target tracking method under low light conditions includes steps S101 to S104, which are described in detail as follows:

[0074] S101, acquiring images frame by frame, identifying targets in the images through a target detector, and determining the target position and target feature information of the current frame;

[0075] This step is fundamental to the entire tracking method. In low-light conditions, image quality is typically poor, with noise and low contrast, making it difficult to accurately identify small objects. The object detector in this step, by introducing a content-aware feature reconstruction (CARAFE) module and a small object detection layer, can enhance the extraction of key target information in low light or complex backgrounds.

[0076] In some embodiments, the image is passed through an object detector to identify the object, and the current frame object position and object feature information are determined, specifically including:

[0077] The object detector includes a perceptual feature reconstruction module and a small object detection layer;

[0078] Determine the target position in the current frame through the perception feature reconstruction module;

[0079] The target feature information is determined through the small target detection layer.

[0080] The Content-Aware Feature Recombination (CARAFE) module dynamically reorganizes feature maps to expand the receptive field, enabling the object detector to capture richer contextual information, thereby improving the recognition of small objects. It can better extract key features of the target in low-light conditions and reduce feature loss caused by insufficient lighting.

[0081] The small object detection layer extracts the fourth downsampling scale from the second layer of the backbone network and fuses feature maps of four different resolutions using FPN and PANet, preserving multi-scale information. This helps accurately detect small objects in low-light images, even when detailed features of these objects are lost due to multiple downsampling operations, thereby improving object detection accuracy.

[0082] In some embodiments, determining the current frame target position by the perceptual feature reconstruction module specifically includes:

[0083] Obtain the feature map corresponding to the target, and perform channel compression on the feature map through 1×1 convolution to generate a compressed feature map;

[0084] Determine the upsampling kernel of a specific size based on the compressed feature map;

[0085] Normalize the upsampling kernel of a specific size to obtain the normalized upsampling kernel;

[0086] Perform element-by-element dot multiplication on the normalized up-sampling kernel and the feature map corresponding to the target to determine the target position of the current frame.

[0087] In some embodiments, as Figure 2As shown in Figure 2, content-aware feature reconstruction (CARAFE) is mainly divided into two steps: upsampling kernel prediction and feature map reconstruction:

[0088] (1) Assuming the upsampling ratio is σ, the input feature map of H×W×C is first compressed by 1×1 convolution to generate H×W×C m The feature map of C m ≤C, reducing the computational complexity of subsequent steps while ensuring effective expression of feature information;

[0089] (2) For the compressed H×W×C m The feature map of encoder ×k encoder The convolution layer is used to predict the upsampling kernel, and the number of input channels is C m , the number of output channels is Expand the channel dimension in the space dimension and get the shape Upsampling kernel of ;

[0090] (3) The obtained up-sampled kernel is normalized using the softmax function so that the sum of the convolution kernel weights is 1;

[0091] (4) Use the upsampling kernel to perform element-by-element dot multiplication with the input feature map to reconstruct the features of each spatial position.

[0092] Through the above steps (1) to (4), the information of each position of the feature map is obtained based on the up and down weighting, which can improve the quality of the reorganized feature map.

[0093] In some embodiments, determining target feature information through the small target detection layer specifically includes:

[0094] Four feature maps of different resolutions of the target are extracted through the small target detection layer, where feature maps of different resolutions represent semantic information at different levels;

[0095] Through the feature pyramid network and path aggregation network, four feature maps of different resolutions are fused to obtain multi-scale information;

[0096] Convolution kernels of different sizes and types are used to perform convolution operations on multi-scale information to obtain target feature information.

[0097] like Figure 3The following figure shows the network structure corresponding to the small object detection layer. In the backbone network, the second layer introduces a fourth scale for downsampling. By combining FPN (Feature Pyramid Network) and PANet (Path Aggregation Network), feature maps of four different resolutions are fused to preserve multi-scale information. A small object detection layer is added to the prediction layer for small-sized object detection. Combined with the other three prediction layers, the prediction layers with different receptive fields are used to further optimize the accuracy of small object detection.

[0098] S102, processing the target position and target feature information of the current frame based on the En-GhostNet network to obtain the target appearance features;

[0099] The main purpose of this step is to extract the appearance features of the target and provide accurate feature representation for subsequent target tracking. The En-GhostNet network has efficient feature learning capabilities and can accurately capture the visual information of the target under low-light conditions.

[0100] Ghost module: This module generates additional feature maps using simple linear operations, reducing the computational overhead of traditional convolution operations. In low-light environments, target features can be more blurred and complex. The Ghost module can reduce computational costs while retaining sufficient feature information, improving feature extraction efficiency.

[0101] Ghost bottleneck architecture: Two Ghost modules are stacked, balancing feature representation capability and computational complexity by adjusting the number of channels and downsampling operations. This architecture enables the En-GhostNet network to better adapt to the feature extraction needs of small objects in low-light conditions while maintaining high computational efficiency.

[0102] In some embodiments, the target position and target feature information of the current frame are processed based on the En-GhostNet network to obtain the target appearance features, specifically including:

[0103] The En-GhostNet network includes Ghost modules and Ghost bottleneck structures;

[0104] The Ghost module generates a corresponding feature map based on the current frame target position and target feature information;

[0105] Build a Ghost bottleneck structure, which includes the first Ghost layer and the second Ghost layer;

[0106] The first Ghost layer expands the number of channels through 1×1 convolution to obtain multi-dimensional detail information of the corresponding feature map;

[0107] The second Ghost layer compresses the number of channels through 1×1 convolution and converts multi-dimensional detail information into target appearance features.

[0108] In some embodiments, for an input image X∈R c×h×w , c represents the number of input image channels, h and w represent the height and width of the input image. The process of generating m feature maps using convolution operation is shown in the following formula:

[0109] Y=X*f

[0110] Among them, * represents the convolution operation, Y∈R h′×w′×m Represents the output feature map of m channels, f∈R c×k×k×n represents the convolution filter of this layer, h' and w' are the height and width of the output image, and k×k is the convolution kernel size of the convolution filter f.

[0111] A series of simple linear operations are applied to each intrinsic feature in Y to obtain the final n feature maps:

[0112]

[0113] Among them, y′ i Represents the i-th intrinsic feature map in Y, and Φ in the above function i Represents the y′ i Get the feature map y ij Linear operations.

[0114] like Figure 4 The following is a schematic diagram of the Ghost module structure: Figure 5 Figure 2 shows a schematic diagram of the Ghost bottleneck structure, which consists of multiple convolutional layers and directly connected layers, primarily implemented by stacking two Ghost modules. The first Ghost module uses 1×1 convolution to expand the number of channels to enhance feature representation and capture more detailed information. The second Ghost module uses 1×1 convolution to compress the number of channels, reducing computational complexity and improving network efficiency. By optimizing the number of channels, the network can extract richer features while maintaining a low computational load. The first Ghost module is followed by a batch normalization (BN) layer and a ReLU activation function, while the second Ghost module is not followed by a ReLU activation. In some cases, a depthwise convolution with a stride of 2 is inserted between the output of the first Ghost module and the input of the second Ghost module to downsample the feature map.

[0115] Specifically, the En-GhostNet feature extraction network adopts the GhostNet architecture, whose core building block is the Ghost bottleneck. It primarily uses the Ghost module to achieve efficient feature extraction. Leveraging the advantages of 1×1 convolution, En-GhostNet can effectively reduce computational complexity while retaining rich feature information, thereby improving network efficiency and performance.

[0116] S103, using Kalman filtering to predict the target position in the next frame corresponding to the target appearance feature;

[0117] The Kalman filter is an optimal estimation algorithm that predicts the next position of a target based on its motion model and current observations. In low-light conditions, the target's motion can be affected by factors such as lighting variations and occlusions, leading to unstable tracking.

[0118] Motion Prediction: The Kalman filter builds a motion model for the target and uses its position and velocity in the current frame to predict its position in the next frame. This helps to reasonably estimate the target's position even when the target is temporarily obscured or illumination changes lead to inaccurate detection, thus improving tracking stability.

[0119] Error Correction: During the prediction process, the Kalman filter corrects the predicted value based on the current detection results, making the prediction more accurate. By continuously updating the target's state estimate, the Kalman filter can adapt to changes in the target's motion and reduce tracking errors caused by lighting and environmental factors.

[0120] In some embodiments, using Kalman filtering to predict the next frame target position corresponding to the target appearance feature specifically includes:

[0121] Determine the target state x based on the target appearance features k , target state x k Including the target position x k and speed v k :

[0122]

[0123] Predict the next frame state of the target according to the following formula:

[0124]

[0125] in, is the target state predicted by the current frame, F is the state transfer matrix, which indicates the change of the target state from the previous frame to the current frame, x k-1 is the target state of the previous frame;

[0126] The Euclidean distance between the target location and the predicted location is calculated according to the following formula:

[0127]

[0128] Among them, (x pred ,y pred ) is the predicted target position, (x det ,y det ) is the detected target position;

[0129] If the Euclidean distance between the target position and the predicted position meets the threshold range, the target position is updated using the current predicted position to determine the target position for the next frame:

[0130] x k =x pred +α·(x det -x pred )

[0131] Among them, x pred is the predicted position, x det is the current detection position, and α is the weight factor.

[0132] Specifically, if the calculated distance meets the threshold range, it means that the detection box and the prediction box are matched successfully, where α is the weight factor, which represents the weighted ratio of the prediction and observation during the update, usually 0≤α≤1.

[0133] At each time step, after prediction, matching, and updating, the target position of the current frame is output as the latest estimate. This position will be used for prediction and updating of the next frame.

[0134] S104: Output target tracking information frame by frame.

[0135] This step is the final output link of the entire tracking method, which organizes and outputs the tracked target information to facilitate subsequent application and analysis.

[0136] Information Completeness: The output tracking information includes the target's bounding box, category label, confidence value, and corresponding tracking ID, providing comprehensive information about the target. This information helps users accurately understand the target's location, type, and tracking status, meeting the needs of different application scenarios.

[0137] Continuous Tracking: Frame-by-frame tracking information ensures continuous tracking and identification of targets throughout the entire video sequence. By assigning a unique tracking ID to each target, it accurately associates the same target across frames, enabling continuous tracking and providing the foundation for subsequent tasks such as behavioral analysis and event detection.

[0138] In some embodiments, outputting target tracking information frame by frame specifically includes:

[0139] The target tracking information includes the target's bounding box information, confidence value, and corresponding tracking ID;

[0140] Determine the corresponding bounding box information according to the target position of the target in the current frame;

[0141] Assign a unique tracking ID to each target based on its location;

[0142] The confidence value is determined based on the Euclidean distance between the target location and the predicted location.

[0143] Specifically, in each frame, the target's bounding box position (x, y, w, h), category label, and confidence score are output. Based on the target detection results, each target is assigned a unique tracking ID and continuously tracked by matching the predicted and detected boxes. The target position is updated in each frame, ensuring that each target is consistently identified and tracked throughout the video sequence.

[0144] This method uses a detection-based small target tracking algorithm, employing a content-aware feature reconstruction (CARAFE) module and a detection layer specifically designed for small targets. This effectively improves the accuracy and robustness of target detection in low-light environments and enhances target recognition in complex environments. Furthermore, the En-GhostNet feature extraction network is proposed to accurately extract the target's appearance features, enabling high-precision target matching in subsequent frames.

[0145] Compared with traditional methods, the present invention not only improves tracking accuracy, but also significantly reduces the consumption of computing resources, improves the overall efficiency of the system, and can better meet the needs of complex application scenarios such as all-weather traffic monitoring and nighttime autonomous driving.

[0146] This application also proposes a small target tracking system under low light conditions, the system comprising: a target detection unit, a feature extraction unit, a prediction unit and an output unit;

[0147] The target detection unit is used to acquire images frame by frame, identify the target in the image through the target detector, and determine the target position and target feature information of the current frame;

[0148] The feature extraction unit processes the target position and target feature information of the current frame based on the En-GhostNet network to obtain the target appearance features;

[0149] The prediction unit uses Kalman filtering to predict the target position in the next frame corresponding to the target appearance features;

[0150] Output unit, outputs target tracking information frame by frame.

[0151] The present application also provides a readable storage medium storing a computer program. When the computer program is executed by a processor, the processor performs the following steps:

[0152] Acquire images frame by frame, identify the target through the target detector, and determine the target position and target feature information of the current frame;

[0153] Based on the En-GhostNet network, the target position and target feature information of the current frame are processed to obtain the target appearance features;

[0154] Use Kalman filtering to predict the target position in the next frame corresponding to the target appearance features;

[0155] Output target tracking information frame by frame.

[0156] The present application also proposes a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer program is executed by the processor to perform the following steps:

[0157] Acquire images frame by frame, identify the target through the target detector, and determine the target position and target feature information of the current frame;

[0158] Based on the En-GhostNet network, the target position and target feature information of the current frame are processed to obtain the target appearance features;

[0159] Use Kalman filtering to predict the target position in the next frame corresponding to the target appearance features;

[0160] Output target tracking information frame by frame.

[0161] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0162] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0163] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person skilled in the art would be able to make numerous modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. The above disclosures are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Therefore, equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A small target tracking method under low light conditions, characterized in that: The method comprises: Acquire images frame by frame, identify targets through target detectors, and determine the target position and target feature information of the current frame; Processing the target position and target feature information of the current frame based on the En-GhostNet network to obtain target appearance features; Using Kalman filtering to predict the target position in the next frame corresponding to the target appearance feature; Output the tracking information of the target frame by frame.

2. The small target tracking method under low light conditions according to claim 1, characterized in that The process of identifying a target in the image through a target detector and determining a target position and target feature information in a current frame specifically includes: The target detector includes a perception feature recombination module and a small target detection layer; Determine the current frame target position by the perception feature reconstruction module; Target feature information is determined through the small target detection layer.

3. The small target tracking method under low light conditions according to claim 2, characterized in that: The determining of the current frame target position by the perception feature recombination module specifically includes: Obtain the feature map corresponding to the target, and perform channel compression on the feature map through 1×1 convolution to generate a compressed feature map; Determine an upsampling kernel of a specific size according to the compressed feature map; Normalizing the upsampling kernel of the specific size to obtain a normalized upsampling kernel; Perform element-by-element dot multiplication on the normalized up-sampling kernel and the feature map corresponding to the target to determine the target position of the current frame.

4. The small target tracking method under low light conditions according to claim 2, characterized in that: The determining target feature information through the small target detection layer specifically includes: Extracting four feature maps of the target at different resolutions through the small target detection layer, wherein the feature maps at different resolutions represent semantic information at different levels; The four feature maps with different resolutions are fused through the feature pyramid network and the path aggregation network to obtain multi-scale information; Convolution kernels of different sizes and types are used to perform convolution operations on multi-scale information to obtain target feature information.

5. The small target tracking method under low light conditions according to claim 1, characterized in that: The processing of the target position and target feature information of the current frame based on the En-GhostNet network to obtain target appearance features specifically includes: The En-GhostNet network includes a Ghost module and a Ghost bottleneck structure; The Ghost module generates a corresponding feature map based on the current frame target position and target feature information; Constructing a Ghost bottleneck structure, wherein the Ghost bottleneck structure includes a first Ghost layer and a second Ghost layer; The first Ghost layer expands the number of channels through 1×1 convolution to obtain multi-dimensional detail information of the corresponding feature map; The second Ghost layer compresses the number of channels through 1×1 convolution and converts the multi-dimensional detail information into target appearance features.

6. The small target tracking method under low light conditions according to claim 1, characterized in that: The method of using Kalman filtering to predict the target position of the next frame corresponding to the target appearance feature specifically includes: Determine the target state x according to the target appearance feature k , the target state x k Including the target position x k and speed v k : Predict the next frame state of the target according to the following formula: in, is the target state predicted by the current frame, F is the state transfer matrix, which indicates the change of the target state from the previous frame to the current frame, x k-1 is the target state of the previous frame; The Euclidean distance between the target position and the predicted position is calculated according to the following formula: Among them, (x pred ,y pred ) is the predicted target position, (x det ,y det ) is the detected target position; If the Euclidean distance between the target position and the predicted position meets the threshold range, the target position is updated using the current predicted position to determine the target position for the next frame: x k =x pred +α·(x det -x pred ) Among them, x pred is the predicted position, xd et is the current detection position, and α is the weight factor.

7. The small target tracking method under low light conditions according to claim 6, characterized in that: Outputting the tracking information of the target frame by frame specifically includes: The tracking information of the target includes the target's bounding box information, confidence value and corresponding tracking ID; Determine the corresponding bounding box information according to the target position of the target in the current frame; Assign a unique tracking ID to each target based on its location; A confidence value is determined based on the Euclidean distance between the target position and the predicted position.

8. A small target tracking system under low light conditions, characterized in that: The system includes: a target detection unit, a feature extraction unit, a prediction unit and an output unit; The target detection unit is used to acquire images frame by frame, identify the target in the image through the target detector, and determine the target position and target feature information of the current frame; A feature extraction unit processes the target position and target feature information of the current frame based on the En-GhostNet network to obtain target appearance features; A prediction unit, using Kalman filtering to predict the target position in the next frame corresponding to the target appearance feature; The output unit outputs the tracking information of the target frame by frame.

9. A readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 7.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 7.

Images