Multi-target floating garbage tracking method based on D-StrongSORT

Abstract

The invention discloses a multi-target water floating garbage tracking method based on D-StrongSORT (Dempster-StrongSORT). Relates to the technical field of image recognition and tracking. A D-StrongSORT multi-target tracking algorithm combining an appearance model and a motion model is used for tracking floating garbage, a previous detection result video is used as input, a previous position state is associated with a next position state by calculating motion features and appearance features, prediction and tracking of a target position can be well achieved, DIOU is adopted by the D-StrongSORT to replace IOU, and the prediction and tracking efficiency is improved. According to the method, a new feature updating strategy is used to match appearance features of a tracking target to reduce ID switching times of the tracking target, adaptive Kalman filtering is used to reduce error accumulation of a motion model, a new feature updating strategy is used to match the appearance features of the tracking target when an appearance model is established, and better correlation matching of the motion model and the appearance model is realized by designing adaptive weighted cost matrix calculation.

Description

Technical Field

[0001] The invention relates to the technical field of video image recognition and processing, in particular to a multi-target floating garbage tracking method based on D-StrongSORT. Background Art

[0002] Multi-Object Tracking (MOT) is a key research direction in the field of computer vision. It aims to detect the positions of multiple targets in real time and accurately in a video sequence, and assign a unique identifier to each target, thereby constructing a complete motion trajectory in consecutive frames. In MOT, the similarity matrix is ​​used to detect the positions of multiple targets in real time and accurately and help construct a complete motion trajectory, and the appearance features of the target are one of the important bases for calculating the similarity matrix.

[0003] The Deep Affinity Network (DAN) proposed by Sun et al. successfully models the appearance of objects in non-adjacent frames and can accurately evaluate their similarity. Yu et al. designed a feature extraction network based on GoogLeNet, which is specifically used to extract the appearance features of the target and embed these features into the tracker, thus achieving performance comparable to that of advanced offline trackers.

[0004] In addition to appearance features, motion features also play an important role in MOT. The SORT algorithm assumes that the displacement of the target between frames follows linear uniform motion, and uses the target position of the previous frame to predict the position of the current frame through the Kalman filter. The OC-SORT algorithm proposed by Cao et al. introduces an observation-centric update strategy, which further reduces the cumulative error and improves the tracking accuracy.

[0005] Appearance models and motion models often complement each other in terms of performance: appearance models perform well when processing complex motion patterns, but their models are usually large and easily affected by external factors such as lighting and occlusion; relatively speaking, motion models are faster and less dependent on detectors, but may show a decrease in tracking capabilities when dealing with complex trajectories. Therefore, multi-target tracking technology that combines the advantages of appearance models and motion models can provide strong support for achieving more accurate and robust monitoring of floating garbage on the water surface.

[0006] DeepSORT further enhances the use of target appearance features based on the SORT algorithm by introducing the re-identification technology in deep learning, thereby improving the robustness and accuracy of tracking. StrongSORT further optimizes DeepSORT and significantly improves the performance of multi-target tracking through more sophisticated algorithms and strategies. However, StrongSORT still uses IOU as the association indicator in the floating garbage target matching stage, but IOU only considers the overlap between the prediction box and the detection box, and fails to fully reflect the actual distance between the two, thus affecting the matching accuracy.

[0007] Tracking floating garbage is a dynamic process. Floating garbage will drift with the river. In this process, the spatial morphology of each floating garbage is highly similar, and the size will continue to change with the impact of the water flow. At the same time, the phenomenon of overlapping and occluding each other by floating garbage often occurs. In addition, the movement trajectory of floating garbage is complex and changeable. In response to the above problems, this invention combines the D-StrongSORT multi-target tracking algorithm of the appearance model and the motion model to realize the tracking of floating garbage. The previous detection result video is used as input. By calculating the motion features and appearance features, the previous position state is associated with the next position state, which can well realize the prediction and tracking of the target position. D-StrongSORT uses DIOU instead of IOU to reduce the number of ID switching of the tracking target, and uses adaptive Kalman filtering to reduce the error accumulation of the motion model. When the appearance model is established, a new feature update strategy is used to match the appearance features of the tracking target. By designing an adaptive weighted cost matrix calculation, a better association matching of the motion model and the appearance model is achieved. Through experiments, it is verified that the algorithm can well complete the task of tracking floating garbage, and the tracking effect is good. Summary of the invention

[0008] The purpose of the present invention is to provide a multi-target floating garbage tracking method based on D-StrongSORT to solve the problems raised in the above background technology.

[0009] To achieve the above object, the present invention provides the following technical solutions:

[0010] A multi-target floating garbage tracking method based on D-StrongSORT includes the following contents: in order to better match the tracking results, appearance information and motion information are used simultaneously in the tracking and matching process of the multi-target tracking model D-StrongSORT, and a cost association mode of the motion model and the appearance model is constructed. The motion model first uses ECC to compensate the camera to reduce the problem of excessive noise scale caused by camera shaking, and then uses adaptive Kalman filtering to predict the future position of the target, associates it with the detected position, and calculates the motion association cost of the target; the appearance model uses ResNeSt50 to extract more target appearance features, uses the EMA feature update strategy to speed up the matching speed and reduce the matching time, and finally calculates the appearance cost of the target according to the detection result, and performs weighted fusion of the appearance cost and the motion cost to output the matching result;

[0011] Appearance Model

[0012] D-StrongSORT adopts the EMA feature update strategy to replace the direct matching from the feature library. It also uses a more powerful appearance feature extractor BoT to replace the simple CNN network. It uses ResNeSt50 as the backbone and pre-trains on the DukeMTMC-reID dataset to extract more appearance features.

[0013] The core idea of ​​the EMA algorithm is to use exponential decay to smooth historical data. In the optimization algorithm, EMA mainly uses the update speed of the adjustment parameters to make it more adaptable to the changing trend of the current data;

[0014]

[0015] in, represents the current time step, It is a smoothing factor, which controls the influence of historical data on the current EMA value. is the current data value, is the EMA value of the previous time step. Through continuous iterative calculation, EMA can smooth historical data and adapt to new data changes more quickly.

[0016] Motion Model

[0017] D-StrongSORT inherits the StrongSORT framework and introduces the NSA-Kalman algorithm and the ECC algorithm. The algorithm automatically adjusts the numerical parameters in the Kalman filter algorithm, such as the covariance matrix of the measurement noise and the prediction noise, to adapt to different environmental conditions. This adaptive adjustment mechanism makes the filter algorithm more applicable in dynamic environments and can more accurately process real-time changing data.

[0018] In the StrongSORT algorithm, the NSA-Kalman algorithm and the ECC algorithm work together to construct a comprehensive motion model. During the target tracking process, the NSA-Kalman algorithm is first used to predict the target position and update the state, and then the ECC algorithm is used to calibrate the target's appearance, ultimately achieving accurate tracking of the target.

[0019] Furthermore, compared with the ResNet network, ResNeSt50 splits the input features according to the number of channels, making full use of the parallel structure to speed up the network's reasoning speed. At the same time, it uses split attention to establish the correlation between the target in channels and space during feature fusion, thereby improving the reliability of the appearance model.

[0020] Furthermore, D-StrongSORT also combines the DIOU distance metric.

[0021] Furthermore, the ECC algorithm is essentially an objective function:

[0022] (1)

[0023] in, and They represent two images with little difference in content but affected by transformations such as lighting, scale, color, and translation.

[0024] (2)

[0025] Among them, warp represents the affine transformation of the image.

[0026] By repeatedly iterating to optimize the objective function, we can solve the equation (4.8) and The corresponding transformation matrix.

[0027] Finally, the D-StrongSORT algorithm uses DIOU in the water drift garbage target matching stage to reduce the number of ID switching times of the tracked target and achieve better target tracking effect.

[0028] (3)

[0029] in, is the Euclidean distance between the center point of the predicted box and the detection box of the floating garbage target, It is the diagonal length of the minimum enclosing rectangular box of the prediction box and detection box of the floating garbage target. is the original intersection-union ratio.

[0030] (4)

[0031] in, is the intersection of the real box and the predicted box, It is the union of the true box and the predicted box.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] 1. Design cost association mode: D-StrongSORT multi-target tracking algorithm combines appearance model and motion model to track floating garbage. The motion model uses ECC camera compensation and NSA Kalman filter, and the appearance model uses ResNeSt50 feature extractor and EMA feature update strategy. D-StrongSORT uses DIOU instead of IOU, which can accurately capture the motion trajectory and appearance characteristics of floating garbage while reducing the number of ID switching of the tracking target. Adaptive Kalman filtering is used to reduce the error accumulation of the motion model.

[0034] 2. Adaptive weighted cost matrix calculation: When the appearance model is established, a new feature update strategy is used to match the appearance features of the tracked target. By designing an adaptive weighted cost matrix calculation, better correlation matching between the motion model and the appearance model can be achieved.

[0035] 3. The motion model uses the NSA-Kalman algorithm and the ECC algorithm: it can adapt to different environmental conditions. This adaptive adjustment mechanism makes the filtering algorithm more applicable in dynamic environments and can more accurately process real-time changing data. The NSA-Kalman algorithm and the ECC algorithm in StrongSORT complement each other and jointly improve the accuracy and stability of target tracking. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is a flow chart of the D-StrongSORT algorithm of the present invention;

[0037] Figure 2 This is a structural diagram of the appearance model of the present invention;

[0038] Figure 3 The Kalman prediction tracking process of the present invention;

[0039] Figure 4 is the distribution of large and small targets in the YRDG dataset of the present invention;

[0040] Figure 5 The training loss and Top1Error curves of the present invention;

[0041] Figure 6 This is a diagram of the floating garbage tracking results of the present invention. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions 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 of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] The present invention is specifically implemented through the following technical solutions:

[0044] 1 Multi-target floating garbage tracking using D-StrongSORT

[0045] The multi-target tracking model D-StrongSORT uses both appearance information and motion information in the tracking and matching process to achieve better matching tracking results. The present invention constructs a cost association model for the motion model and the appearance model. The motion model first uses ECC to compensate the camera to reduce the problem of excessive noise scale caused by camera shake, and then uses an adaptive Kalman filter to predict the future position of the target, associate it with the detected position, and calculate the motion association cost of the target; the appearance model uses ResNeSt50 to extract more target appearance features, and uses the EMA feature update strategy to speed up the matching speed and reduce the matching time. Finally, based on the detection results, the appearance cost of the target is calculated. The appearance cost and motion cost are weightedly fused to output the matching result. The model structure is as follows: Figure 1 shown.

[0046] 2 Appearance model establishment

[0047] D-StrongSORT adopts the EMA feature update strategy to replace the direct matching from the feature library, and uses a more powerful appearance feature extractor BoT instead of a simple CNN network. ResNeSt50 is used as the backbone and pre-trained on the DukeMTMC-reID dataset to extract more appearance features.

[0048] The core idea of ​​the EMA algorithm is to use exponential decay to smooth historical data. In the optimization algorithm, EMA mainly uses the update speed of the adjustment parameters to make it more adaptable to the changing trend of the current data.

[0049]

[0050] in, represents the current time step, It is a smoothing factor, which controls the influence of historical data on the current EMA value. is the current data value, is the EMA value of the previous time step. Through continuous iterative calculation, EMA can smooth historical data and adapt to new data changes more quickly. The appearance model structure is as follows Figure 2 shown.

[0051] Compared with the ResNet network, ResNeSt50 divides the input features according to the number of channels, makes full use of the parallel structure, and speeds up the network's reasoning speed. At the same time, it uses split attention to establish the correlation between the target in channels and space during feature fusion, thereby improving the reliability of the appearance model.

[0052] 3 Motion model construction

[0053] The motion model uses the NSA-Kalman algorithm and the ECC algorithm. The Kalman filter algorithm is a classic target tracking algorithm, but it also has certain limitations and is only applicable to linear prediction. D-StrongSORT has made key innovations based on the inheritance of the StrongSORT framework. The NSA-Kalman algorithm was introduced, which automatically adjusts the numerical parameters in the Kalman filter algorithm, such as the covariance matrix of the measurement noise and the prediction noise, to adapt to different environmental conditions. This adaptive adjustment mechanism makes the filter algorithm more applicable in dynamic environments and can more accurately process real-time changing data. The NSA-Kalman algorithm and the ECC algorithm in StrongSORT complement each other and jointly improve the accuracy and stability of target tracking. In particular, D-StrongSORT also combines the DIOU distance metric to further improve the accuracy of target matching. It consists of a state estimation step and a state update step. The Kalman prediction tracking process is as follows Figure 3 shown.

[0054] Each target has its corresponding trajectory. For each trajectory, the successfully associated target trajectory is calculated. When the Kalman filter predicts, it increases with the successful association of the video, and the tracking ends when the maximum threshold (that is, the maximum number of frames) is reached. For targets that cannot be associated with existing trajectories, the system considers the target to be a new tracking target, and then starts a new trajectory hypothesis, that is, if these new trajectories can be successfully associated in 3 consecutive frames, they are determined to be newly appeared tracking targets; if they are not successfully associated in 3 consecutive frames, the hypothesis is not true and the trajectory is deleted.

[0055] The NSA-Kalman algorithm provides a basic target tracking framework that uses the Kalman filter to perform state estimation and update to achieve target position prediction and tracking. The algorithm performs well in processing linear predictions, but may have limited effectiveness in the face of nonlinear transformations or large appearance differences.

[0056] To solve these problems, StrongSORT introduces the ECC algorithm to enhance the correlation coefficient, thereby improving the accuracy of data association between targets. Through the ECC algorithm, the appearance of the target between different frames can be calibrated during the tracking process, so that the target can be matched more accurately even when there are changes such as lighting, scale, color, and translation.

[0057] In the StrongSORT algorithm, the NSA-Kalman algorithm and the ECC algorithm work together to build a comprehensive motion model. In the process of target tracking, the NSA-Kalman algorithm is first used to predict the target position and update the state, and then the ECC algorithm is used to calibrate the target's appearance, and finally the target is accurately tracked.

[0058] The essence of the ECC algorithm is an objective function:

[0059] (1)

[0060] in, and They represent two images with little difference in content but affected by transformations such as lighting, scale, color, and translation.

[0061] (2)

[0062] Among them, warp represents the affine transformation of the image.

[0063] By repeatedly iterating to optimize the objective function, we can solve the equation (4.8) and The corresponding transformation matrix.

[0064] Finally, the D-StrongSORT algorithm uses DIOU instead of IOU in the floating garbage target matching stage to reduce the number of ID switching times of the tracked target and achieve better target tracking effect.

[0065] (3)

[0066] in, is the Euclidean distance between the center point of the predicted box and the detection box of the floating garbage target, It is the diagonal length of the minimum enclosing rectangular box of the prediction box and detection box of the floating garbage target. is the original intersection-union ratio.

[0067] (4)

[0068] in, is the intersection of the real box and the predicted box, It is the union of the true box and the predicted box.

[0069] Experiment and result analysis

[0070] The dataset of the present invention is collected from the Yellow River floating garbage. The dataset mainly consists of inland river scenes, with a total of 2400 images. Multi-target accumulated garbage accounts for a large proportion of the images, and is often accompanied by unknown floating garbage, which is difficult to distinguish.

[0071] like Figure 4 As shown, in the floating garbage dataset constructed by the present invention, small targets account for more than 84% of the proportion. Not only that, in various categories, small targets also occupy a dominant position. This data feature makes the dataset uniquely valuable in research and application. Therefore, the dataset can be used to test and optimize small target and multi-target detection and tracking algorithms, which helps to improve the accuracy and robustness of the algorithm when dealing with small targets such as floating garbage in actual scenarios. At the same time, the dataset also provides researchers in related fields with rich experimental data, which helps to promote the development and application of floating garbage detection and tracking technology.

[0072] The experiment uses the deep learning framework Pytorch, Python3.7.6, CUDA10.0, and the GPU is NVIDIA GeForceRTX 3060 for single-card training. The experimental parameters of the D-StrongSORT tracking algorithm are set as follows: the number of initial frames for tracking the target is 3 frames, the maximum threshold distance for IOU matching is 0.6; the maximum number of frames retained for lost tracking is 30; the number of frames for appearance feature calculation N=100. The target detection model uses the above target detection algorithm model YOLOv7 to ensure the fastest detection speed and the highest detection accuracy; the image size of the input model is 640×640; the detection box confidence threshold is 0.3 to ensure a sufficient number of detection boxes; the IOU threshold for NMS non-maximum screening in YOLOv7 is 0.5.

[0073] In order to measure the superiority of the algorithm in this complex dataset, the tracking accuracy (MOT Accuracy, MOTA), tracking precision (MOT Precision, MOTP), ID switch times (ID switch, IDsw) and most successfully tracked trajectories (Mostly Tracked Trajectories, MT) are selected.

[0074] (1)

[0075] in, is the index of the time frame, is the target of the real frame, (False Negative) is the number of real target objects not detected by this method (missed detection). (False Positive) is the number of objects that are falsely detected by the method but do not exist in the real target object. IDSW is defined as an IDSW that occurs when an object is occluded and then reappears in the field of view. If the network reassigns a sequence number, it is defined as a IDSW.

[0076] (2)

[0077] in, Indicates Frame Target and assumptions The number of matches; For the Frame Target An example of the position between it and its paired hypothesis, i.e. the matching error.

[0078] MT is the number of trajectories where most targets are tracked, that is, the number of correct target object trajectories where the number of frames successfully tracked accounts for more than 80% of the total number of frames.

[0079] Comparative experimental analysis

[0080] Table 1 intuitively shows the performance differences of different tracking algorithms in the floating garbage tracking task.

[0081] Table 1 Differences between different algorithms

[0082]

[0083] The SORT algorithm has a MOTA of 57.18%, indicating that it has low performance in dealing with false positives, false negatives, and identity switches. With the iteration and optimization of the algorithm, DeepSORT improved to 76.04%, StrongSORT further improved to 79.63%, and D-StrongSORT reached the highest 81.17%. This continuous improvement shows the significant progress of the algorithm in maintaining tracking stability and accuracy. Multi-target tracking accuracy measures the spatial accuracy of tracking. SORT has a MOTP of 75.40%, while D-StrongSORT's MOTP increased to 86.37%, with each algorithm iteration bringing an improvement in the accuracy of position prediction. This improvement may be due to more optimized feature extraction and target localization techniques. The number of identity switches decreased from 39 in SORT to 15 in D-StrongSORT, indicating that the algorithm's ability to maintain the target identity during tracking has been significantly enhanced, which is particularly important in dynamic and complex tracking environments and helps reduce false positives and tracking confusion. The number of multiple tracked targets increased from 49 in SORT to 79 in D-StrongSORT, reflecting the improvement in the algorithm’s ability to continuously track the same target. This improvement indicates that the algorithm is more effective in managing the appearance and disappearance of targets in the video sequence, enhancing the continuity and reliability of long-term tracking.

[0084] Experimental results show that D-StrongSORT is superior to other algorithms in terms of evaluation indicators such as tracking accuracy, tracking precision, number of ID switching and successful tracking trajectory, and can effectively avoid missed tracking and ID exchange phenomena.

[0085] Training results analysis

[0086] When D-StrongSORT builds the appearance model, it uses ResNeSt50 as the backbone network, so it is necessary to train the network to generate feature weights for small floating garbage targets. After extracting the features of floating garbage of different appearance types through ResNeSt50, they are used to calculate the cosine similarity of the appearance model.

[0087] The training loss curve and Top1Error curve are as follows Figure 5 As shown in the figure, the training loss curve represents the loss during the training process, the lower the better. The Top1Error curve means: assuming that the model predicts the category of an object and the model outputs 1 prediction result, the probability that this result can be judged correctly is the Top-1 accuracy rate. The probability of misjudgment is the Top-1 error rate, which is simply the probability of the model making a wrong judgment. The smaller the loss curve and the error curve, the better the model performance. Figure 5It can be seen that with the increase in the number of training times, the loss and error curves become smaller and smaller, and finally tend to be stable, indicating that the network is becoming more and more stable and accurate.

[0088] Visualization of experimental analysis

[0089] The visualization results are as follows Figure 6 As shown. The comparison of (a1)(a2) shows that the SORT algorithm has missed tracking and ID exchange, and there is an error in tracking the overlapping target in the (a1) picture; the comparison of (b1)(b2) shows that the DeepSORT algorithm has produced two ID exchanges, which needs to be optimized; the comparison of (c1)(c2) shows that the tracked floating garbage target has an ID switch, and according to the comparison of (c2)(c3), it is shown that the tracking of multiple tracking targets has no ID switch. The algorithm D-StrongSORT proposed in the present invention does not have ID exchange and missed tracking phenomena in (d1), (d2) and (d3), and the recognition probability is guaranteed, indicating that the tracking effect of the method proposed in the present invention is good and can meet the basic requirements in the field of floating garbage tracking.

[0090] It is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting from any point of view, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention, and any reference numerals in the claims should not be regarded as limiting the claims involved.

[0091] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A multi-target floating garbage tracking method based on D-StrongSORT, characterized in that: Includes the following: In order to better match the tracking results, the appearance information and motion information are used simultaneously in the tracking and matching process of the multi-target tracking model D-StrongSORT, and the cost association mode of the motion model and the appearance model is constructed. The motion model first uses the Enhanced Correlation Coefficient (ECC) algorithm to compensate the camera to reduce the problem of excessive noise scale caused by camera shake, and then uses the adaptive Kalman filter to predict the future position of the target, associate it with the detected position, and calculate the motion association cost of the target; the appearance model uses ResNeSt50 to extract more target appearance features, and uses the Exponential Moving Average (EMA) feature update strategy to speed up the matching speed and reduce the matching time. Finally, according to the detection results, the appearance cost of the target is calculated, and the appearance cost and motion cost are weightedly fused to output the matching result; Appearance model: D-StrongSORT adopts the EMA feature update strategy to replace the direct matching from the feature library. It also uses a more powerful appearance feature extractor BoT (Bag of Tricks) instead of a simple CNN network. It uses ResNeSt50 as the backbone and pre-trains on the DukeMTMC-reID dataset to extract more appearance features. The core idea of ​​the EMA algorithm is to use exponential decay to smooth historical data. In the optimization algorithm, EMA mainly uses the update speed of the adjustment parameters to make it more adaptable to the changing trend of the current data; in, represents the current time step, It is a smoothing factor, which controls the influence of historical data on the current EMA value. is the current data value, It is the EMA value of the previous time step. Through continuous iterative calculation, EMA can smooth historical data and adapt to new data changes more quickly. Sports Model: D-StrongSORT inherits the StrongSORT framework and introduces the NoiseScale Adaptive Kalman (NSA-Kalman) algorithm and the ECC algorithm. The algorithm automatically adjusts the numerical parameters in the Kalman filter algorithm, such as the covariance matrix of the measurement noise and the prediction noise, to adapt to different environmental conditions. This adaptive adjustment mechanism makes the filter algorithm more applicable in dynamic environments and can more accurately process real-time changing data. In the StrongSORT algorithm, the NSA-Kalman algorithm and the ECC algorithm work together to construct a comprehensive motion model. During the target tracking process, the NSA-Kalman algorithm is first used to predict the target position and update the state, and then the ECC algorithm is used to calibrate the target's appearance, ultimately achieving accurate tracking of the target.

2. The multi-target floating garbage tracking method based on D-StrongSORT according to claim 1 is characterized by: Compared with the ResNet network, ResNeSt50 divides the input features according to the number of channels, makes full use of the parallel structure, and speeds up the network's reasoning speed. At the same time, it uses split attention to establish the correlation between the target in channels and space during feature fusion, thereby improving the reliability of the appearance model.

3. The multi-target floating garbage tracking method based on D-StrongSORT according to claim 1 is characterized by: D-StrongSORT also incorporates the DIOU distance metric.

4. The multi-target floating garbage tracking method based on D-StrongSORT according to claim 3 is characterized by: The ECC algorithm is essentially an objective function: (1) in, and They represent two images with little content difference, but affected by transformations such as lighting, scale, color, and translation; (2) Among them, warp represents the affine transformation of the image; By repeatedly iterating to optimize the objective function, we can solve the equation (4.8) and The corresponding transformation matrix; Finally, the D-StrongSORT algorithm uses DIOU in the floating garbage target matching stage to reduce the number of ID switching times of the tracked target and achieve better target tracking effect; (3) in, is the Euclidean distance between the center point of the predicted box and the detection box of the floating garbage target, The diagonal length of the minimum circumscribed rectangular box of the prediction box and detection box of the floating garbage target; is the original intersection ratio; (4) in, is the intersection of the real box and the predicted box, It is the union of the true box and the predicted box.

Images