Target tracking method, electronic equipment and computer readable storage medium

Abstract

The invention provides a target tracking method, electronic equipment and a computer readable storage medium. According to the target tracking method provided by the invention, through combination of fusion processing of the target detection model and the motion model, the tracking precision of the infrared small target in the satellite remote sensing image is effectively improved. By acquiring the first position information of the target in each frame and performing prediction processing in combination with the position information of the previous frame, the method can correct the position of the target in real time, and reduces interference caused by a complex background and a low signal-to-noise ratio. Particularly, when the target is shielded or moves, the position of the target can be accurately corrected by fusing detection and prediction information, so that the robustness and accuracy of tracking are improved, and the requirement of high-precision tracking is met.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a target tracking method, an electronic device, and a computer-readable storage medium. Background Art

[0002] Infrared small target tracking refers to the technology of detecting and locating targets based on the thermal radiation characteristics of the targets in infrared images or videos.

[0003] In the scenario of infrared small target tracking based on satellite remote sensing images, there is a technical problem of low tracking accuracy. Summary of the Invention

[0004] The purpose of the present invention is to overcome the above-mentioned technical deficiencies and provide a target tracking method, an electronic device and a computer-readable storage medium to solve the technical problem of low tracking accuracy in the related art in the scenario of infrared small target tracking based on satellite remote sensing images.

[0005] In order to achieve the above technical objectives, the present invention adopts the following technical solutions: In a first aspect, the present invention provides a target tracking method, comprising: Receiving satellite remote sensing image data of a current frame; wherein the satellite remote sensing image data carries at least one target; Inputting the satellite remote sensing image data into a preset target detection model to obtain first position information of the target in the current frame; The second position information of a predetermined target is input into a preset motion model for prediction processing to obtain third position information, and then the third position information and the first position information are fused to obtain fourth position information; wherein, the second position information is the final position information of the target in the previous frame, the third position information is the position information of the current frame predicted by the motion model based on the second position information, and the fourth position information is the final position information of the target in the current frame.

[0006] Furthermore, the preset target detection model includes: A feature extraction network is used to perform feature extraction on satellite remote sensing image data to obtain a feature map; A region proposal network is used to generate first position information based on the feature map.

[0007] Furthermore, the feature extraction network is also configured to retain a feature map with a target resolution through a preset convolutional neural network layer structure when performing feature extraction, and introduce an expanded convolution operation in at least one convolution layer to expand the receptive field.

[0008] Furthermore, the feature extraction network is a pre-improved ResNet-50, which includes: The first Stage layer is used to perform downsampling on the input image to obtain a low-level feature map; The second stage layer is used to perform downsampling on the low-level feature map to obtain the mid-level feature map; The third stage layer is used to expand the receptive field through a preset dilated convolution operation while maintaining the resolution of the intermediate feature map to obtain a feature map with a target resolution.

[0009] Furthermore, the target detection model is pre-trained using a preset total loss function, and the total loss function includes a first loss function and a second loss function; wherein, the first loss function is a classification loss function, and the second loss function is an offset loss function; wherein, the classification loss function is used to characterize the degree of matching between the category confidence output by the target detection model and the true annotation, and the offset loss function is used to characterize the regression error between the target bounding box coordinate prediction value and the true value.

[0010] Furthermore, the preset motion model is a preset uniform motion model, a preset accelerated motion model, a preset adaptive filtering model, or a prediction model obtained through training based on historical data.

[0011] Furthermore, the preset motion model is a preset Kalman filter model; the step of inputting the second position information of the predetermined target into the preset motion model for prediction processing to obtain third position information, and then fusing the third position information with the first position information to obtain fourth position information includes: Processing the second position information based on a state transition parameter preset in a Kalman filter model to generate third position information; Based on the state transition parameters, the covariance matrix corresponding to the previous frame, and the process noise covariance matrix, a priori covariance matrix corresponding to the current frame is generated; wherein the covariance matrix corresponding to the previous frame is used to characterize the quantization of uncertainty of the target state estimate in the previous frame, and the priori covariance matrix corresponding to the current frame is represented as a quantization of uncertainty of the target state estimate in the predicted current frame; Calculating a Kalman gain matrix based on a priori covariance matrix corresponding to the current frame, a preset measurement matrix, and an observation noise covariance matrix; The Kalman gain matrix, the first position information, the observation matrix, and the third position information are input into a preset fusion function for processing to obtain the fourth position information.

[0012] Furthermore, the process noise covariance matrix is ​​updated using the following method: Calculating movement change data of the target based on the first position information and the second position information; Determining whether the movement change data is greater than a preset change threshold; When the variance is greater than a preset threshold, the variance corresponding to the process noise covariance matrix is ​​increased to enhance the adaptability to non-uniform motion.

[0013] In a second aspect, the present invention provides an electronic device comprising: a memory, and one or more processors communicatively connected to the memory; the memory stores instructions executable by the one or more processors, and the instructions are executed by the one or more processors to enable the one or more processors to implement the above-mentioned method.

[0014] In a third aspect, the present invention provides a computer-readable storage medium, wherein the computer program is stored in the computer-readable storage medium, and the computer program implements the above method when executed by a processor.

[0015] Beneficial effects: The target tracking method provided by this invention effectively improves the tracking accuracy of small infrared targets in satellite remote sensing images by integrating a target detection model with a motion model. By acquiring the target's initial position information in each frame and performing a predictive process based on the position information from the previous frame, this method can correct the target's position in real time, reducing interference caused by complex backgrounds and low signal-to-noise ratios. In particular, when the target is obscured or experiences motion changes, the fusion of detection and prediction information allows for precise correction of the target's position, thereby improving the robustness and accuracy of tracking and meeting the requirements for high-precision tracking. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is one of the flow diagrams of a target tracking method provided by an embodiment of the present invention; Figure 2 This is a second flow chart of a target tracking method provided by an embodiment of the present invention; Figure 3 : is a schematic diagram of the effect of small target tracking in the prior art provided by an embodiment of the present invention; wherein, Figure 3 The left picture in the figure is a schematic diagram of the effect of tracking a small target at the previous moment. Figure 3 The right picture in the figure is a schematic diagram of the effect of tracking a small target at the next moment; Figure 4 : is a schematic diagram of the effect of using the present embodiment of the present invention to track a small target; wherein, Figure 4 The left picture in the figure is a schematic diagram of the effect of tracking a small target at the previous moment. Figure 4 The right picture in the figure is a schematic diagram of the effect of tracking a small target at the next moment; Figure 5 This is a block diagram of an electronic device used in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] In order to enable those skilled in the art to better understand the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention 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, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of this application.

[0018] In the field of infrared small target tracking based on satellite remote sensing imagery, with the development of satellite remote sensing technology, using satellite remote sensing imagery to track targets has become an important application. However, due to the special properties of satellite remote sensing images, target tracking accuracy faces significant challenges in many cases, especially for tracking small infrared targets.

[0019] In the related technologies, in the field of infrared small target tracking for satellite earth observation, existing technologies face many defects, which lead to serious limitations in target tracking accuracy, especially in complex backgrounds and low signal-to-noise ratio environments.

[0020] On the one hand, small infrared targets often only occupy a few pixels in the image and lack recognizable shape and texture features, making them extremely difficult to effectively distinguish using traditional image processing methods. In one possible implementation, by performing multi-level downsampling operations, the effective activation area of ​​the target may be compressed to 1-2 pixels. Moreover, due to the complete loss of thermal radiation gradient details at the edge of the target and insufficient distinction from noise points under low signal-to-noise ratio conditions, the confidence level of target detection for small targets is often too low. Therefore, this makes the effective feature extraction of small infrared targets a huge bottleneck, seriously affecting the subsequent tracking accuracy.

[0021] On the other hand, satellite remote sensing images often include complex backgrounds, such as thermal radiation reflected from urban buildings or solar flares on the ocean surface, which often results in a very small grayscale difference between the target and the background. In one possible implementation, the Region Proposal Network (RPN) is often designed for anchor boxes of 16×16 pixels or larger. Such large anchor boxes are often unable to effectively handle very small objects of several pixels. Therefore, in low signal-to-noise ratio scenarios, the false detection rate is often too high. The influence of background interference further weakens the ability to distinguish between the target and the background, making the detection results of small objects unstable and affecting subsequent trajectory prediction.

[0022] Furthermore, small infrared targets in satellite imagery may appear to be in constant or maneuvering motion. In one possible implementation, trajectory prediction relies on a single constant-motion model. However, when the target's scale changes rapidly, the existing motion model cannot adaptively adjust noise parameters, leading to cumulative position prediction errors and excessive trajectory drift. This is particularly true in scenarios where the target undergoes rapid maneuvers or changes, as traditional constant-motion models are unable to adequately address these issues, resulting in tracking failure.

[0023] Understandably, in the Siamese networks used in related technologies, excessive downsampling of the backbone network causes the morphological features of small targets to disappear, making it more difficult to identify targets in images. The motion models and target detection results in related technologies lack a closed-loop mechanism for bidirectional correction. This prevents the algorithm from fully leveraging the relationship between historical position information and real-time detection results, resulting in issues with tracking accuracy and stability.

[0024] It is also understandable that the three aforementioned flaws are interrelated, forming a cascading failure chain. Specifically, insufficient feature extraction leads to reduced confidence in target detection, and thus a decrease in the reliability of the observations fed into the motion model. This further amplifies the prediction errors of the fixed-parameter motion model, leading to increased trajectory drift and ultimately a breakdown in tracking accuracy. When the target is obscured by clouds or undergoes a maneuver, the methods in the related art lack an effective closed-loop correction mechanism, resulting in an excessively high tracking interruption rate.

[0025] In summary, in the scenario of infrared small target tracking based on satellite remote sensing images, there is a technical problem of low tracking accuracy.

[0026] This embodiment provides a target tracking method. The execution subject of this embodiment can be a deep learning server, an embedded processing unit, an edge computing device, a computing platform of a satellite ground station, etc.

[0027] like Figure 1 and Figure 2 As shown, the method may include: Step S12: receiving satellite remote sensing image data of the current frame; wherein the satellite remote sensing image data carries at least one target.

[0028] In this embodiment, the current frame may represent a piece of image data received at the current time point during target tracking (real-time tracking). During satellite remote sensing image processing, images are often acquired continuously in a time series, with each frame representing the satellite's observation of the ground at a specific moment or over a period of time.

[0029] In this embodiment, the current frame can also represent the satellite remote sensing image data currently being processed by the execution entity (e.g., a deep learning server). In other words, another device (computing platform) can pre-acquire multiple frames of satellite remote sensing image data and sequentially transmit these frames to the deep learning server for algorithm processing to obtain the target's location information (non-real-time tracking).

[0030] In this embodiment, the satellite remote sensing image data may be represented as ground image information collected by a remote sensing sensor (eg, an infrared sensor, an optical sensor, a radar sensor, etc.) carried by a satellite.

[0031] In this embodiment, the satellite remote sensing image data may be infrared image data. Specifically, it will be understood that infrared images acquire image data by sensing the thermal radiation of ground targets. Unlike visible light images, infrared images do not rely on external light sources, but rather on the thermal radiation of the target itself or the background. Therefore, infrared images can still function effectively in low-light environments (such as at night or under cloud cover) and can identify targets on the ground.

[0032] Understandably, infrared image data often has a low signal-to-noise ratio (SNR). This is especially true when ground temperature variations are minimal or the background is complex, where the temperature difference between the target and background in the infrared image can be very small. With a low SNR, the grayscale difference between the target and background is minimal, making it difficult to discern the edges and details of the target, thus increasing the difficulty of target detection.

[0033] Understandably, targets in satellite remote sensing infrared imagery are often small. Especially in high-resolution images, small infrared targets may occupy only 1-10 pixels. This makes the target's morphological features relatively inconspicuous and its difference from the background less pronounced, increasing the complexity of target detection and tracking. In particular, in low signal-to-noise ratio environments, the target's thermal radiation details may be severely affected, further reducing the confidence level of target detection.

[0034] Understandably, infrared image data is often affected by complex backgrounds. For example, thermal radiation reflected from urban buildings, solar flares on the ocean surface, and cloud cover can interfere with target recognition. The background's temperature distribution can be very similar to that of the target, making the target's thermal radiation signal less distinct in the image and increasing confusion between the target and the background.

[0035] In this embodiment, the target may be a ground vehicle, an aircraft, or a human being, etc.

[0036] Step S14: inputting the satellite remote sensing image data into a preset target detection model to obtain first position information of the target in the current frame.

[0037] In this embodiment, the preset target detection model can be a twin network feature extraction module. Specifically, the twin network feature extraction module can use an improved ResNet-50 (Residual Network-50) backbone network + Region Proposal Network (RPN) as the overall architecture.

[0038] In this embodiment, the preset object detection model can also be an object detection model based on the YOLO (You Only Look Once) algorithm. It is understood that YOLO is an end-to-end object detection model with efficient real-time performance and strong object detection capabilities. YOLO can simultaneously perform object classification and bounding box regression in a single forward propagation, thereby achieving rapid object detection.

[0039] In this embodiment, the preset target detection model can also be a target detection model based on the Faster R-CNN (Region-based Convolutional Neural Network) algorithm. It is understood that the target detection architecture of Faster R-CNN combines the Region Proposal Network (RPN) and the Convolutional Neural Network (CNN). By generating candidate regions and performing classification and regression, it can effectively identify and locate targets in images.

[0040] In this embodiment, the preset target detection model may also be a target detection model based on the SSD (Single Shot MultiBox Detector) algorithm.

[0041] In this embodiment, the preset target detection model may also be a target detection model based on EfficientDet.

[0042] In this embodiment, the preset target detection model may also be a target detection model based on the Transformer architecture.

[0043] In this embodiment, the first position information may include the position information (x, y, w, h) of the candidate frame with the highest confidence, and may also include the offset (Δx, Δy, Δw, Δh) of the candidate frame.

[0044] Step S16: Input the second position information of the predetermined target into a preset motion model for prediction processing to obtain third position information, and then fuse the third position information and the first position information to obtain fourth position information; wherein, the second position information is the final position information of the target in the previous frame, the third position information is the position information of the current frame predicted by the motion model based on the second position information, and the fourth position information is the final position information of the target in the current frame.

[0045] In this embodiment, the second position information represents the target's final position in the previous frame. It will be appreciated that the method of this embodiment can be applied to multiple frames of satellite remote sensing image data in a time series (one frame of satellite remote sensing image data corresponding to each time point). For each frame of satellite remote sensing image data in the time series, the method of this embodiment can determine the final position information corresponding to that frame (or time point). In other words, the target's final position information for a particular frame or time point serves as the input information for calculating the final position information for the next frame or time point.

[0046] Specifically, the second position information reflects the target's exact location in the previous frame. This information is derived from the combination of the target detection model and the motion model to determine the target's coordinates in the previous frame. This position information represents the target's state at the previous moment in the time series, including parameters such as its position, size, and shape in the image. This position has been corrected using the fusion of target detection and motion modeling, providing accurate target state information.

[0047] In this embodiment, the second position information serves as input data, providing the known state of the target in the previous frame. This information is then fed into the motion model to predict the target's position in the current frame (i.e., the third position information). This second position information provides information about the target's historical motion, allowing subsequent predictions to more accurately infer the target's position at the next moment, rather than relying solely on the target detection results in the current frame.

[0048] In this embodiment, the third position information is the target position information obtained by using a motion model to predict the target's final position information in the previous frame (i.e., the second position information). The motion model can be a uniform motion model, an acceleration model, a Kalman filter model, a particle filter model, or the like, and can predict the target's possible position in the current frame based on the target's dynamic characteristics.

[0049] In this embodiment, the third position information may include the target's coordinates (predicted, not final) and the target's motion state information in the current frame (predicted, not final), such as velocity and acceleration. It should be understood that the third position information is solely predicted by the motion model. However, to improve tracking accuracy and robustness, it is necessary to combine the predicted results with those of the target detection model. This allows the system to utilize historical position information to correct inaccuracies in detection results, thereby improving the reliability and accuracy of the prediction.

[0050] In this embodiment, the preset motion model may be a preset uniform motion model, a preset accelerated motion model, a preset adaptive filtering model, or a prediction model trained based on historical data. Specifically, the second position information may be input into the preset uniform motion model, the preset accelerated motion model, the preset adaptive filtering model, or the prediction model trained based on historical data for calculation, thereby obtaining the third position information.

[0051] In this embodiment, the uniform motion model can assume that the target's velocity remains constant within each time step. Therefore, the target's position in the current frame is predicted based on the target's position and velocity in the previous frame. In this model, the target's motion is linear, and the target moves at a constant velocity over time. Therefore, the target's position in the current frame can be inferred from the target's position and velocity in the previous frame, assuming that the target's velocity remains unchanged. The position and velocity of the previous frame are then combined to predict the position of the current frame.

[0052] In this embodiment, the acceleration motion model assumes that the speed of the target changes within each time step, that is, the target has a certain acceleration. The prediction of the third position information can be calculated based on the position information, speed and acceleration of the target in the previous frame. The acceleration model needs to consider the impact of the target's acceleration on the position. Therefore, the movement of the target in each frame is not only determined by the speed, but also needs to be calculated based on the acceleration of the previous frame. It can be understood that in the acceleration motion model, the movement of the target is nonlinear, and the target position and speed changes in the current frame will be affected by the target acceleration.

[0053] In this embodiment, the adaptive filtering model can be an adaptive filtering model based on Kalman filtering, or it can be a particle filtering model or a Gaussian process regression model, etc. The adaptive filtering model can dynamically update the target position prediction and adaptively adjust the noise parameters in the prediction process. The third position information can be predicted by the target position and speed of the previous frame, as well as the estimated covariance. The adaptive filtering model updates the motion model of the target at each time step based on the actual measurement and prediction results. In this way, the prediction of the third position information is not only based on the historical target position, but also takes into account the uncertainty and noise factors of the target motion, so the prediction can be dynamically adjusted to adapt to different motion modes, especially when the target changes speed or is subject to external interference.

[0054] In this embodiment, the prediction model trained based on historical data can be a prediction model based on a deep learning algorithm, or a prediction model based on a machine learning algorithm. Specifically, the prediction model based on the deep learning algorithm can be a long short-term memory network (LSTM), a recurrent neural network (RNN), or a deep reinforcement learning model, etc. The prediction model based on the machine learning algorithm can be a decision tree regression model, a random forest regression model, or a linear regression model, etc. The historical data can be the target's position information, speed information, acceleration information, and other feature data related to the target's motion in the previous time frame. These historical data can be used to train the prediction model so that the prediction model can predict the target's future position by learning its motion patterns.

[0055] The target tracking method provided in this embodiment effectively improves the tracking accuracy of small infrared targets in satellite remote sensing images by combining a target detection model with a motion model. By obtaining the target's first position information in each frame and performing predictive processing based on the position information of the previous frame, this method can correct the target's position in real time, reducing interference caused by complex backgrounds and low signal-to-noise ratios. In particular, when the target is obscured or changes motion, the fusion of detection and prediction information can achieve precise correction of the target's position, thereby improving the robustness and accuracy of tracking and meeting the requirements of high-precision tracking.

[0056] In some embodiments, the preset target detection model includes: The feature extraction network is used to perform feature extraction on satellite remote sensing image data to obtain a feature map.

[0057] In this embodiment, the feature extraction network may be a pre-improved ResNet-50, which is configured to retain a feature map with a target resolution through a preset convolutional neural network layer structure when performing feature extraction, and introduce an expanded convolution operation in at least one convolutional layer to expand the receptive field.

[0058] In this embodiment, the feature extraction network can be a modified VGG-16 (Visual Geometry Group). By removing the last two max pooling layers, the network's output feature map resolution can be increased from the traditional 7×7 to 14×14 pixels. This allows the network to maintain high feature expression capabilities without reducing resolution, making it suitable for small object detection tasks.

[0059] In this embodiment, the feature extraction network may be an improved MobileNetV3 (MobileNetwork Version 3). The depthwise separable convolution stride of the final stage (Stage 5) may be changed from 2 to 1 to maintain a higher resolution.

[0060] A region proposal network is used to generate first position information based on the feature map.

[0061] In this implementation, the Region Proposal Network (RPN) is used to generate candidate regions for subsequent object detection and classification. Specifically, the RPN generates a set of anchor boxes based on the feature map of the input image and predicts whether each anchor box contains an object and the precise location of the anchor box (bounding box adjustment).

[0062] Specifically, at each location in the feature map, the RPN generates multiple anchor boxes (which can be rectangular boxes of different scales and aspect ratios). These anchor boxes are used to cover areas where objects may appear. Each anchor box is a preliminary guess about the region that may or may not contain an object. The size and scale of the anchor box are predefined to ensure that objects of different sizes and shapes in the image can be covered. For each anchor box, the RPN performs a binary classification: determining whether the anchor box contains the object (target box) or whether it is background (non-target box). This process can be implemented using a binary classifier, which outputs the probability of whether each anchor box contains an object.

[0063] RPN not only predicts whether an anchor box contains an object, but also performs bounding box regression on the anchor box, adjusting the four bounding parameters of the anchor box (i.e., x, y position, width, and height) to more accurately fit the actual location of the object. The goal of regression is to make the anchor box fit the object's position more closely, thereby improving the accuracy of subsequent object positioning.

[0064] In a specific implementation scheme, the preset target detection model adopts an improved ResNet-50 backbone network + region proposal network (RPN) as the overall architecture, which is aimed at the detection and tracking tasks of small infrared targets ≤10×10 pixels in satellite scenarios.

[0065] For the improved ResNet-50 backbone network, the Stage 4 structure was deleted, while the Stage 1 and Stage 2 structures were retained. All downsampling operations were canceled in Stage 3 (the convolution stride was uniformly set to 1), the feature map resolution was increased to 28×28, and the number of residual blocks and output channels was increased (256) to enhance feature representation capabilities. In addition, a dilated convolution structure with a dilation rate of 2 was introduced in Stage 3, effectively expanding the receptive field to 7×7 pixels (the corresponding receptive area of ​​the original image was 56×56 pixels), enhancing the modeling capability of the surrounding radiation gradient.

[0066] In this specific implementation scheme, candidate box generation and sample control are achieved through the RPN mechanism: three very small anchor boxes are set (corresponding to 16×16, 24×24, and 32×32 pixels of the original image), and FocalLoss is used to perform weighted optimization on the sample imbalance problem, while combining the SmoothL1 loss to regress the coordinate offset of the candidate box.

[0067] The overall architecture outputs three key pieces of information: a high-resolution feature map (28×28×256) that serves as the input for subsequent modules; the position information (x, y, w, h) of the candidate boxes generated by the RPN; and the confidence values ​​and offsets (Δx, Δy, Δw, Δh) of the candidate boxes. The classification branch outputs confidence scores for target presence determination, while the regression branch outputs offsets for precise positioning. Through structural optimization and scale adaptation, this module achieves significant response to faint infrared targets, providing high-quality initial information support for subsequent trajectory prediction and tracking head modules.

[0068] This embodiment effectively improves the accuracy and efficiency of target detection by adopting a preset target detection model, which includes a feature extraction network and a region candidate network. The feature extraction network can perform deep feature extraction on satellite remote sensing image data to obtain rich feature maps that can capture the key details and texture information of the target. The region candidate network generates candidate regions based on these feature maps, and then generates the first position information of the target. Through this method, the target can be accurately located, and even under complex backgrounds and low signal-to-noise ratio conditions, false detection and missed detection can be effectively reduced, thereby improving the robustness and accuracy of target detection. Therefore, this embodiment provides reliable technical support for tracking small targets in satellite remote sensing images, which helps to improve the overall performance and stability of target tracking.

[0069] In some embodiments, the feature extraction network is further configured to retain a feature map with a target resolution through a preset convolutional neural network layer structure when performing feature extraction, and introduce an expanded convolution operation in at least one convolutional layer to expand the receptive field.

[0070] In this embodiment, the preset convolutional neural network layer structure can be a pre-improved ResNet-50, which is configured to retain feature maps with target resolution through the preset convolutional neural network layer structure when performing feature extraction, and introduce an expanded convolution operation in at least one convolution layer to expand the receptive field.

[0071] In this embodiment, the preset convolutional neural network layer structure can be a modified VGG-16. By removing the last two max pooling layers, the network's output feature map resolution can be increased from the traditional 7×7 to 14×14 pixels. This allows the network to maintain high feature expression capabilities without reducing resolution, making it suitable for small object detection tasks.

[0072] In this embodiment, the preset convolutional neural network layer structure can be an improved MobileNetV3. The depthwise separable convolution step size of the final stage (Stage 5) can be changed from 2 to 1 to maintain a higher resolution.

[0073] In this embodiment, target resolution can be expressed as the spatial detail preservation capability required by the feature map output by the feature extraction network. This ensures that the target's thermal radiation gradient details are effectively preserved, thereby supporting the subsequent detection module's sub-pixel positioning accuracy for ultra-small targets (1-10 pixels). In other words, the core requirement for target resolution is to ensure that small targets occupy sufficient activation area in the feature map so that their details are not compressed or lost.

[0074] In this embodiment, the target resolution may be an absolute resolution. For example, the feature map size is ≥ 1 / 8 of the input image size (e.g., input 224×224 → output ≥ 28×28 pixels). This prevents excessive compression of target features in the feature map, preventing the target size from being reduced to an excessively small area of ​​1-2 pixels.

[0075] In this embodiment, the dilated convolution operation can be expressed as inserting holes (i.e., blank areas) between the elements of the convolution kernel to expand it without increasing the amount of computation or reducing the resolution. Specifically, dilated convolution adjusts the dilation rate of the convolution kernel so that the convolution operation can cover a larger area while maintaining a low computational cost. In other words, by inserting holes to expand the receptive field, the network can obtain more contextual information, which can effectively improve the accuracy of target detection, especially when dealing with tasks with complex backgrounds and small targets.

[0076] This embodiment can effectively expand the receptive field while retaining the feature map of the target resolution by introducing the dilated convolution operation in the feature extraction network. This design enables the network to cover a wider range of contextual information when extracting image features, and enhances the perception of small targets and edge targets. By adjusting the receptive field of the convolution kernel, the dilated convolution can capture more global information without increasing the amount of computation, effectively reducing the impact of background interference. At the same time, by retaining the high-resolution feature map, the network can maintain the accuracy of details in target detection, especially under low signal-to-noise ratio or complex background conditions, significantly improving the robustness and accuracy of detection. The introduction of this method improves the network's ability to recognize targets of different scales, shapes and motion states, and enhances the effectiveness and efficiency of the target detection system in practical applications.

[0077] In some embodiments, the feature extraction network is a pre-improved ResNet-50, which includes: The first stage layer is used to perform downsampling on the input image to obtain a low-level feature map.

[0078] In this embodiment, the first stage layer is used to perform preliminary downsampling on the input image and extract low-level features of the image. Specifically, this layer may include a 7×7 convolution kernel (stride 2) and a 3×3 maximum pooling layer (stride 2), followed by three residual blocks. In this layer, a 224×224 pixel infrared image is input, processed by a 7×7 convolution kernel to output a 112×112 pixel feature map, and then processed by a maximum pooling layer to output a 56×56 pixel low-level feature map. The residual block performs a channel expansion operation on the low-level feature map, outputting a 56×56×256 feature map to the second stage layer.

[0079] The second stage layer is used to perform downsampling on the low-level feature map to obtain the mid-level feature map.

[0080] In this implementation, the second stage layer processes the low-level feature maps output by stage 1, continues downsampling, and extracts deeper semantic features. This second stage layer primarily increases the abstraction level of target features in the image through additional convolution operations, helping the network better understand the target's shape and structure.

[0081] In this embodiment, this layer may include 4 residual blocks, where the first residual block has a stride of 2 and the subsequent residual blocks have a stride of 1. In this layer, a 56×56×256 feature map is input, and the first residual block downsamples to output a 28×28 pixel feature map. The subsequent residual blocks maintain the 28×28 resolution and increase the number of channels to 512, outputting a 28×28×512 intermediate feature map to the third stage layer.

[0082] The third stage layer is used to expand the receptive field through a preset dilated convolution operation while maintaining the resolution of the intermediate feature map to obtain a feature map with a target resolution.

[0083] In this implementation, the third stage layer eliminates traditional downsampling and introduces dilated convolution to expand the receptive field while maintaining high feature map resolution. This design aims to enhance the network's ability to perceive small objects, particularly infrared targets (1-10 pixels in size).

[0084] In this embodiment, this layer may include 4 residual blocks, all of which have a fixed stride of 1 and a set number of output channels of 256. A dilated convolution operation with a dilation rate of 2 is integrated into the 3×3 convolution layer of the first residual block. In this layer, a 28×28×512 feature map is input, and the resolution is maintained at 28×28 after processing by a residual block with a stride of 1. The dilated convolution operation doubles the convolution kernel sampling interval, and outputs a 28×28×256 high-resolution feature map to the region proposal network.

[0085] This implementation method can effectively expand the receptive field while maintaining high-resolution feature maps by improving ResNet-50, especially introducing dilated convolution operations in the third Stage layer. This design improves the network's performance in processing complex scenes, especially the detection capabilities of small targets and detailed information. By performing dilated convolution operations in the third Stage layer, the network can capture a wider range of contextual information without sacrificing resolution, thereby avoiding the information loss caused by traditional downsampling. Correspondingly, the first and second Stage layers continue to perform conventional downsampling operations to help the network extract low-level and mid-level features, effectively and gradually enhancing the semantic information of the image. This improvement not only improves the network's perception ability, especially under conditions of low resolution and complex backgrounds, but also enhances the accuracy and robustness of target detection, and is particularly suitable for processing tasks involving small targets or relatively complex backgrounds.

[0086] In some embodiments, the target detection model is pre-trained using a preset total loss function, and the total loss function includes a first loss function and a second loss function; wherein the first loss function is a classification loss function, and the second loss function is an offset loss function; wherein the classification loss function is used to characterize the degree of matching between the category confidence output by the target detection model and the true annotation, and the offset loss function is used to characterize the regression error between the target bounding box coordinate prediction value and the true value.

[0087] In this embodiment, the total loss function can be expressed as:

[0088] In the formula, Expressed as the total loss function, is the classification loss function, is the offset loss function, and is the weight coefficient used to balance the impact of classification loss and offset loss.

[0089] In this embodiment, the classification loss function can be expressed as:

[0090] In the formula, Expressed as the negative sample suppression coefficient, it is used to reduce the loss contribution of simple negative samples (background). It is expressed as a tuning parameter in the focal loss function, which is used to control the degree of attention paid to difficult-to-classify samples. It is the category confidence output by the model, which indicates the probability that a target belongs to a category.

[0091] In this embodiment, the offset loss function can be expressed as:

[0092] In the formula, Expressed as L1 loss, that is, calculating the absolute difference between the predicted value and the true value, and represents the offset of the center coordinate of the target bounding box predicted by the model (i.e., the predicted offset), and Expressed as an offset from the center coordinates of the true object bounding box.

[0093] This embodiment significantly improves the accuracy and robustness of target detection by adopting a preset total loss function to train the target detection model. The total loss function consists of two parts: the classification loss function and the offset loss function, which ensures that the model can achieve high accuracy in both category prediction and target positioning. The classification loss function measures the degree of match between the category confidence output by the target detection model and the true annotation, so that the model can more accurately identify the target category, thereby reducing false detection and missed detection. The offset loss function optimizes the accuracy of the target position by calculating the regression error between the coordinate prediction value and the true value of the target bounding box, especially when the target scale varies greatly or the targets overlap, ensuring the accurate positioning of the bounding box. Through the design of this dual loss function, the model can not only improve the accuracy of target classification, but also accurately locate the target boundary in complex scenarios, thereby providing more stable and accurate target detection results in a variety of application scenarios.

[0094] In some embodiments, the preset motion model is a preset uniform motion model, a preset accelerated motion model, a preset adaptive filtering model, or a prediction model obtained through training based on historical data.

[0095] In this embodiment, the preset motion model may be a preset uniform motion model, a preset accelerated motion model, a preset adaptive filtering model, or a prediction model trained based on historical data. Specifically, the second position information may be input into the preset uniform motion model, the preset accelerated motion model, the preset adaptive filtering model, or the prediction model trained based on historical data for calculation, thereby obtaining the third position information.

[0096] In this embodiment, the uniform motion model can assume that the target's velocity remains constant within each time step. Therefore, the target's position in the current frame is predicted based on the target's position and velocity in the previous frame. In this model, the target's motion is linear, and the target moves at a constant velocity over time. Therefore, the target's position in the current frame can be inferred from the target's position and velocity in the previous frame, assuming that the target's velocity remains unchanged. The position and velocity of the previous frame are then combined to predict the position of the current frame.

[0097] In this embodiment, the acceleration motion model assumes that the speed of the target changes within each time step, that is, the target has a certain acceleration. The prediction of the third position information can be calculated based on the position information, speed and acceleration of the target in the previous frame. The acceleration model needs to consider the impact of the target's acceleration on the position. Therefore, the movement of the target in each frame is not only determined by the speed, but also needs to be calculated based on the acceleration of the previous frame. It can be understood that in the acceleration motion model, the movement of the target is nonlinear, and the target position and speed changes in the current frame will be affected by the target acceleration.

[0098] In this embodiment, the adaptive filtering model can be an adaptive filtering model based on Kalman filtering, or it can be a particle filtering model or a Gaussian process regression model, etc. The adaptive filtering model can dynamically update the target position prediction and adaptively adjust the noise parameters in the prediction process. The third position information can be predicted by the target position and speed of the previous frame, as well as the estimated covariance. The adaptive filtering model updates the motion model of the target at each time step based on the actual measurement and prediction results. In this way, the prediction of the third position information is not only based on the historical target position, but also takes into account the uncertainty and noise factors of the target motion, so the prediction can be dynamically adjusted to adapt to different motion modes, especially when the target changes speed or is subject to external interference.

[0099] In this embodiment, the prediction model trained based on historical data can be a prediction model based on a deep learning algorithm or a prediction model based on a machine learning algorithm. Specifically, the prediction model based on a deep learning algorithm can be a long short-term memory network (LSTM), a recurrent neural network (RNN), or a deep reinforcement learning model, among others. The prediction model based on a machine learning algorithm can be a decision tree regression model, a random forest regression model, or a linear regression model, among others. The historical data can be the target's position information, velocity information, acceleration information, and other feature data related to the target's motion in previous time frames. This historical data can be used to train the prediction model, enabling the prediction model to predict the target's future position by learning its motion patterns.

[0100] This embodiment adopts multiple motion models, including a uniform motion model, an accelerated motion model, an adaptive filtering model, and a prediction model trained based on historical data. It can dynamically select the most appropriate model based on different scenarios and target motion characteristics, thereby effectively improving the accuracy and robustness of target tracking. The uniform motion model is suitable for targets with relatively stable motion and can quickly predict the target's trajectory. The accelerated motion model can capture the motion characteristics of the target during acceleration or deceleration, improving the adaptability to rapidly changing targets. The adaptive filtering model further improves the adaptability to targets in complex dynamic environments by adjusting the parameters of the prediction model in real time. The prediction model trained based on historical data can optimize the prediction of the target trajectory by learning past motion patterns. This embodiment can handle various motion patterns and significantly improve the stability and accuracy of target tracking in changing environments. In particular, it can effectively avoid tracking interruptions and error accumulation when facing nonlinear and sudden motion.

[0101] In some embodiments, the preset motion model is a preset Kalman filter model; the step of inputting the predetermined second position information of the target into the preset motion model for prediction processing to obtain third position information, and then fusing the third position information with the first position information to obtain fourth position information includes: Step S162: Processing the second position information based on state transition parameters preset in the Kalman filter model to generate third position information.

[0102] In this implementation, the pre-set Kalman filter model utilizes four-dimensional state space modeling. Through a closed-loop filtering mechanism of "state initialization-prediction-observation update-closed-loop correction," this module achieves stable trajectory estimation for small infrared targets in complex scenarios. Specifically, this module performs data processing including state initialization, state prediction, observation update, and detection-prediction closed-loop correction. It aims to combine visual detection results with target motion models to enhance the continuous tracking capability of weak-signal targets.

[0103] In this embodiment, the four-dimensional state vector of the target can be defined as:

[0104] In the formula, It can represent the initial position of the target in the image coordinate system, which can be provided by manual annotation and represents the center coordinates of the target in the initial frame. It can represent the initial velocity of the target, indicating the initial velocity of the target in the x-axis and y-axis directions. The initial velocity can be set to 0, which means that the target is assumed to be stationary at the beginning.

[0105] In this embodiment, the initial covariance matrix can be defined as:

[0106] In the formula, can represent the initial covariance matrix, Represent the initial error variance of the target position in the x-axis and y-axis directions, Represent the initial error variance of the target velocity in the x-axis and y-axis directions respectively. This initial covariance matrix can quantify the position uncertainty of the manual annotation.

[0107] In this implementation, process noise can be defined as:

[0108] In the formula, It can represent the variance of process noise and the uncertainty in the target state prediction process. is a 4-dimensional identity matrix, which represents the independence of each component (position and velocity) in the state vector and ensures that the process noise variance of each component is independent.

[0109] In this embodiment, the observation noise can be defined as:

[0110] In the formula, Represents observation noise, which has a mean of 0 and a variance of Gaussian distribution.

[0111] In this embodiment, the preset state transition parameters may be:

[0112] In the formula, is the state transition matrix, which represents the dynamic change of the target from one time point to the next time point. Through this matrix, the algorithm can predict the state at the next moment based on the current state. is the time interval, which indicates the time difference from the previous frame to the current frame.

[0113] In this embodiment, the second position information can be processed based on the state transition parameters preset in the Kalman filter model to generate the third position information using the following formula:

[0114] In the formula, It can be expressed as a predicted state vector, which is the predicted position of the current frame (third position information) calculated based on the target state of the previous frame. It can be expressed as the optimal estimate of the predicted state and the observed value in the previous moment or the previous frame (second position information).

[0115] Step S164: Based on the state transfer parameters, the covariance matrix corresponding to the previous frame and the process noise covariance matrix, generate a priori covariance matrix corresponding to the current frame; wherein, the covariance matrix corresponding to the previous frame is used to characterize the uncertainty quantization of the target state estimation in the previous frame, and the priori covariance matrix corresponding to the current frame is expressed as a quantization of the uncertainty of the target state estimation in the predicted current frame.

[0116] In this embodiment, the preset state transition parameters may be:

[0117] In the formula, is the state transition matrix, which represents the dynamic change of the target from one time point to the next time point. Through this matrix, the algorithm can predict the state at the next moment based on the current state. is the time interval, which indicates the time difference from the previous frame to the current frame.

[0118] In this embodiment, the prior covariance matrix corresponding to the current frame can be generated in the following manner:

[0119] In the formula, is the prior covariance matrix corresponding to the current frame, is the process noise covariance matrix, is the covariance matrix corresponding to the previous frame.

[0120] Step S166: Calculate the Kalman gain matrix based on the prior covariance matrix corresponding to the current frame, the preset measurement matrix and the observation noise covariance matrix.

[0121] In this embodiment, the Kalman gain matrix can be calculated as follows:

[0122] In the formula, is the Kalman gain matrix, is the prior covariance matrix corresponding to the current frame, is the preset observation matrix, is the observation noise covariance matrix.

[0123] Step S168: input the Kalman gain matrix, the first position information, the observation matrix, and the third position information into a preset fusion function for processing to obtain fourth position information.

[0124] In this embodiment, the fusion function may be:

[0125] In the formula, is the Kalman gain matrix, It can be expressed as the predicted state vector (third position information), is the fourth position information, It is the first location information.

[0126] In this embodiment, the covariance matrix (posterior covariance matrix) can be generated by the following formula:

[0127] In the formula, is the identity matrix.

[0128] This implementation significantly improves the accuracy and robustness of target tracking by employing a Kalman filter model and combining multiple steps to predict and fuse target position information. First, based on the state transition function in the Kalman filter, the model accurately infers the predicted position for the current frame from the position information of the previous frame, ensuring the continuity and stability of the target trajectory. Then, by calculating the prior covariance matrix, the model quantifies the uncertainty of the target state, thereby more accurately reflecting the potential errors and uncertainties in the target during the prediction process. Next, the Kalman gain matrix is ​​calculated, further optimizing the trade-off between prediction and actual observation, effectively enhancing the ability to correct the target state. Finally, by fusing the Kalman gain matrix with the target's observation information, the predicted and observed information are efficiently integrated, resulting in more accurate target position information. Overall, this implementation is able to handle complex motion patterns and uncertainties in dynamically changing environments, reducing error accumulation and improving target tracking accuracy. It maintains efficient and stable tracking performance, especially in situations where the target's motion speed varies significantly or there are brief occlusions.

[0129] In some embodiments, the process noise covariance matrix is ​​updated using the following method: Calculating movement change data of the target based on the first position information and the second position information; In this embodiment, the target position in the current frame (i.e., the center coordinates obtained by the target detection model) and the target position in the previous frame (i.e., the final corrected position in the previous frame) are first obtained. By comparing the target positions at these two time points, the target displacement change between the two consecutive frames can be calculated. Therefore, in this embodiment, the displacement change data can be the displacement change between the two consecutive frames.

[0130] In this embodiment, the motion change data may also be the change in the diagonal length of the target bounding box between two consecutive frames (the scale change rate). That is, the diagonal length of the target bounding box of the current frame minus the diagonal length of the target bounding box of the previous frame is divided by the diagonal length of the target bounding box of the previous frame.

[0131] Determining whether the movement change data is greater than a preset change threshold; In this embodiment, the change threshold may be a distance threshold or a scale change rate threshold. For example, the preset threshold may be 0.2, that is, if the target scale changes by more than 20%, it is considered that the target has undergone a significant sudden change in motion, which may be non-uniform motion.

[0132] In other words, if the motion change data exceeds the threshold of 0.2 (the scale change rate), the target is judged to have undergone non-uniform maneuvering, such as nonlinear motion such as acceleration and turning. If the motion change data is ≤ 0.2, the target is judged to have maintained uniform or near-uniform motion, that is, the target's movement is relatively stable with little change.

[0133] When the variance is greater than a preset threshold, the variance corresponding to the process noise covariance matrix is ​​increased to enhance the adaptability to non-uniform motion.

[0134] In this embodiment, the variance in the process noise covariance matrix may be increased to a preset upper limit value.

[0135] This implementation significantly improves the adaptability of the target tracking algorithm in complex dynamic environments by dynamically updating the process noise covariance matrix. When the target's motion change data (i.e., displacement or velocity change) exceeds a preset change threshold, the variance of the process noise covariance matrix is ​​automatically increased, thereby enhancing adaptability to the target's non-uniform motion. This mechanism enables the model to more flexibly adjust its prediction strategy when faced with complex motions such as target acceleration, deceleration, or sudden changes in direction, reducing error accumulation and trajectory drift. By dynamically adjusting the process noise, the algorithm can effectively respond to changes in target motion, improving tracking accuracy and robustness. This ensures continuous and stable target tracking, particularly in situations where the target's motion pattern is uncertain or there is strong environmental interference.

[0136] In a specific implementation, the LaSOT large-scale dataset (generalized generalized object features) and the self-built SIRST satellite dataset (136 sequences, 150,185 images, covering complex backgrounds such as cities and oceans) can be jointly trained to improve the model's adaptability to satellite scenes. High-resolution feature maps (28×28 pixels) and dilated convolutions (7×7 receptive field) enhance the ability to capture edge thermal radiation details of 1-10 pixel targets, increasing target detection confidence in low signal-to-noise ratio scenarios.

[0137] Kalman filtering reduces the incidence of trajectory drift and the number of tracking interruptions in scenarios where the target is maneuvering rapidly or is temporarily blocked.

[0138] On the SIRST dataset, this implementation achieved a success rate of 0.513 and an accuracy of 0.818, which are 13.2% and 14.7% higher than the baseline method SiamRPN (Siamese Region Proposal), respectively. This is especially true in scenarios where the target is extremely small (≤8×8 pixels), meeting the requirements of satellite remote sensing for sub-pixel tracking accuracy.

[0139] After each frame of the image outputs the target candidate box through the twin network, the Kalman filter module is triggered: if the detection confidence is ≥0.5, the "detection + prediction" fusion result is adopted; if the confidence is <0.5, the pure trajectory prediction mode is enabled until the target is recaptured.

[0140] This implementation plan provides an efficient solution for tracking small infrared targets in satellite scenarios through the deep integration of "feature enhancement" and "trajectory correction", significantly improving the tracking accuracy and real-time performance in complex environments. In the SIRST dataset test, the algorithm processing speed reached 65 frames per second, meeting the real-time requirements of the satellite platform; in the scenario of rapid target maneuvering (speed 15 pixels / frame), the tracking success rate increased by 22% compared with the baseline method, verifying the effectiveness of "deep feature extraction + spatiotemporal trajectory correction" and having important engineering application value.

[0141] Specific as Figure 3 and Figure 4 As shown, from Figure 3 It can be seen that the existing technology has certain errors when tracking small targets, especially when the target's motion changes greatly, the position deviation of the target frame is obvious, resulting in a decrease in tracking accuracy. This error manifests as the target's bounding box not accurately covering the target, especially when the target is maneuvering or changing rapidly, trajectory drift or false detection may occur. This phenomenon illustrates the limitations of the existing technology in dealing with the rapid motion or direction changes of small targets. It cannot fully adapt to changes in the target state and easily leads to tracking failure. Figure 4 It's clear that using the method of the present invention, the target frame always accurately follows the target's position. Regardless of the target's speed or direction, the present invention can accurately correct the target's trajectory in real time. The target frame remains stably over the target, demonstrating the present invention's precise adaptability to motion conditions when tracking small targets.

[0142] According to an embodiment of the present invention, an electronic device is provided. Figure 5The electronic device in this embodiment may include one or more of the following components: a processor, a network interface, a memory, a non-volatile memory, and one or more applications, wherein the one or more applications may be stored in the non-volatile memory and configured to be executed by one or more processors, and the one or more programs are configured to execute the method described in the aforementioned method embodiment.

[0143] According to an embodiment of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a computer, the computer executes the method described in any one of the above embodiments.

[0144] According to an embodiment of the present invention, a computer program product comprising instructions is further provided. When the instructions are executed by a computer, the computer is enabled to perform a method in any one of the above embodiments.

[0145] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in a sequence other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0146] Optionally, the specific examples in this embodiment may refer to the examples described in the above embodiments, and this embodiment will not be described in detail here.

[0147] The serial numbers of the above embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0148] In the above embodiments of the present application, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0149] The above is only a preferred embodiment of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A target tracking method, characterized in that: include: Receiving satellite remote sensing image data of a current frame; wherein the satellite remote sensing image data carries at least one target; Inputting the satellite remote sensing image data into a preset target detection model to obtain first position information of the target in the current frame; The second position information of a predetermined target is input into a preset motion model for prediction processing to obtain third position information, and then the third position information and the first position information are fused to obtain fourth position information; wherein, the second position information is the final position information of the target in the previous frame, the third position information is the position information of the current frame predicted by the motion model based on the second position information, and the fourth position information is the final position information of the target in the current frame.

2. The method according to claim 1, characterized in that The preset target detection model includes: A feature extraction network is used to perform feature extraction on satellite remote sensing image data to obtain a feature map; A region proposal network is used to generate first position information based on the feature map.

3. The method according to claim 2, characterized in that The feature extraction network is also configured to retain a feature map with a target resolution through a preset convolutional neural network layer structure when performing feature extraction, and introduce an expanded convolution operation in at least one convolution layer to expand the receptive field.

4. The method according to claim 3, characterized in that The feature extraction network is a pre-improved ResNet-50, which includes: The first Stage layer is used to perform downsampling on the input image to obtain a low-level feature map; The second stage layer is used to perform downsampling on the low-level feature map to obtain the mid-level feature map; The third stage layer is used to expand the receptive field through a preset dilated convolution operation while maintaining the resolution of the intermediate feature map to obtain a feature map with a target resolution.

5. The method according to claim 4, characterized in that The target detection model is pre-trained using a preset total loss function, which includes a first loss function and a second loss function; wherein the first loss function is a classification loss function and the second loss function is an offset loss function; wherein the classification loss function is used to characterize the degree of matching between the category confidence output by the target detection model and the true annotation, and the offset loss function is used to characterize the regression error between the target bounding box coordinate prediction value and the true value.

6. The method according to claim 1, characterized in that The preset motion model is a preset uniform motion model, a preset accelerated motion model, a preset adaptive filtering model, or a prediction model obtained through training based on historical data.

7. The method according to claim 1, characterized in that The preset motion model is a preset Kalman filter model; the step of inputting the second position information of the predetermined target into the preset motion model for prediction processing to obtain third position information, and then fusing the third position information with the first position information to obtain fourth position information includes: Processing the second position information based on a state transition parameter preset in a Kalman filter model to generate third position information; Based on the state transition parameters, the covariance matrix corresponding to the previous frame, and the process noise covariance matrix, a priori covariance matrix corresponding to the current frame is generated; wherein the covariance matrix corresponding to the previous frame is used to characterize the quantization of uncertainty of the target state estimate in the previous frame, and the priori covariance matrix corresponding to the current frame is represented as a quantization of uncertainty of the target state estimate in the predicted current frame; Calculating a Kalman gain matrix based on a priori covariance matrix corresponding to the current frame, a preset measurement matrix, and an observation noise covariance matrix; The Kalman gain matrix, the first position information, the observation matrix, and the third position information are input into a preset fusion function for processing to obtain the fourth position information.

8. The method according to claim 6, characterized in that The process noise covariance matrix is ​​updated using the following method: Calculating movement change data of the target based on the first position information and the second position information; Determining whether the movement change data is greater than a preset change threshold; When the variance is greater than a preset threshold, the variance corresponding to the process noise covariance matrix is ​​increased to enhance the adaptability to non-uniform motion.

9. An electronic device, characterized in that: include: a memory, and one or more processors communicatively coupled to the memory; Instructions executable by the one or more processors are stored in the memory. The instructions are executed by the one or more processors to enable the one or more processors to implement the method according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that The readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images