Open vocabulary multi-target tracking method and system for foggy day traffic scene

Abstract

The invention discloses an open vocabulary multi-target tracking method and system for a foggy day traffic scene, and the method comprises the steps: generating a cross-modal target suggestion through a vision-language combined modeling-based open vocabulary detector through a re-parameterized vision-language path aggregation network and region-text comparative learning; the method can flexibly recognize the unlabeled unknown category target in the training set, breaks through the dependence of a traditional detector on a fixed category label, and effectively solves the problems of target appearance feature degradation and background noise interference in a foggy day scene. The traditional NMS algorithm is optimized by combining the distance of the center points of the bounding boxes and the information of the overlapping regions, false detection and missing detection are remarkably reduced, the accuracy of target box screening is improved, and a solid foundation is laid for follow-up track association. A self-adaptive trajectory interpolation method based on Gaussian process regression is introduced, a nonlinear motion mode of a target is captured through a radial basis function kernel, a missing trajectory is dynamically repaired, and continuity and stability of trajectory prediction are enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent transportation, in particular to an open-vocabulary multi-target tracking method for foggy traffic scenes, which is suitable for scenarios such as automatic driving and intelligent monitoring that need to maintain tracking robustness in bad weather. BACKGROUND

[0002] The existing foggy multi-target tracking method faces three major technical bottlenecks, which significantly limits its application effect in complex traffic scenes. First, the traditional method highly depends on the fixed class label library and can only identify the predefined training set class. However, for unknown class targets in the open traffic scene, the miss detection rate is as high as more than 30%. This class dependency makes the model lack generalization ability in diversified and dynamic real environment, and it is difficult to adapt to the demand of flexible target identification of intelligent transportation system. Second, in the scene of target frame overlap, the traditional non-maximum suppression (NMS) algorithm performs poorly. The target feature blur caused by thick fog and the complexity of the background increase the difficulty of distinguishing overlapping targets, and the suppression strategy of traditional NMS based on simple IoU threshold is easy to mistakenly remove the real target frame, which leads to serious missing detection problem and further affects the accuracy of subsequent trajectory association. Finally, the lack of trajectory breakage repair capability is another key bottleneck. The traditional linear interpolation method only relies on the endpoint position of the previous and next frames, and it is difficult to effectively fit the nonlinear motion mode such as turning and speed change, which leads to significant deviation of trajectory prediction. The high frequency of ID switching of traditional method reflects its limitation in maintaining the consistency of target identity in dynamic scene. Although the improvement scheme in recent years tries to alleviate some problems by enhancing the foggy feature extraction, its generalization ability is limited, and in addition, the real-time performance on edge devices is poor and the frame rate is low, which is difficult to meet the demand of high-efficiency inference in automatic driving or intelligent monitoring system. These bottlenecks together highlight the lack of robustness, real-time performance and adaptability of existing methods in foggy environment. SUMMARY

[0003] The purpose of the present application is to provide an open-vocabulary multi-target tracking method for foggy traffic scenes and a system, which is suitable for multi-target tracking in foggy low-visibility scenes, solves the defects of existing multi-target tracking methods in class generalization, target overlap misjudgment and trajectory breakage repair (target feature blur, unknown class and trajectory breakage in foggy environment), and provides an efficient and robust multi-target tracking solution for intelligent transportation systems and automatic driving technology.

[0004] To achieve the above purpose, the technical scheme adopted by the present application is:

[0005] The open-vocabulary multi-target tracking method for foggy traffic scenes includes the following contents:

[0006] Step S1, the input image is processed by the open vocabulary detector of visual-linguistic joint modeling, and the motion features of the detection result are extracted;

[0007] Step S2, the extracted motion features are matched by center distance weighted DIoU measurement, and the matched trajectory is updated;

[0008] Step S3, the motion trajectory is modeled based on Gaussian process regression by Gaussian smoothing interpolation processing, and the smoothing degree is adaptively adjusted by using dynamic scale parameter, to obtain the repaired trajectory;

[0009] Step S4, all trajectories at the current time are obtained according to the matched trajectory and the repaired trajectory, and the matched trajectory at the current time is output.

[0010] The specific processing procedure of the open vocabulary detector is as follows:

[0011] Step 11, the target image input is obtained, and the image quality is optimized by a light fog filtering module;

[0012] Step 12, visual-linguistic feature fusion is performed one by one, and the RepVL-PAN network performs multi-scale fusion of input multi-scale image features and text embedding, text guided feature enhancement, and image pooling attention to update text embedding;

[0013] Step 13, the Prompt-Then-Detect paradigm is adopted, the open vocabulary set is predefined, the prompt word embedding is precomputed offline, and is directly called during inference;

[0014] Step 14, the detection result is output, and the target bounding box and the corresponding embedding feature are generated.

[0015] The specific processing procedure of step S2 is as follows: the motion features of the extracted high-score detection result are matched for the first time, and the motion features of the extracted low-score detection result are matched for the second time.

[0016] The DIoU measurement processing procedure is as follows:

[0017] Step 201, the position deviation between the boxes is quantified by the Euclidean distance square d;

[0018] Step 202, the distance is normalized by the diagonal length c of the minimum circumscribed rectangle;

[0019] Step 303, for the adjacent boxes satisfying DIoU>θ, the highest confidence box is retained, and the rest is set to zero.

[0020] The Gaussian smoothing interpolation processing procedure of step S3 is as follows:

[0021] Step 31, trajectory missing frame modeling: the missing frame position P tConstruct a Gaussian process regression model:

[0022] P t =f (i) (t)+∈#(4-6)

[0023] Wherein, ∈~N(0, σ 2 ) is Gaussian noise, f (i) (t) obeys zero-mean Gaussian process:

[0024] f (i) ~GP(0, k(x, x'))#(4-7);

[0025] Step 32, dynamic kernel adjustment: adopt radial basis kernel function

[0026]

[0027] Wherein, the scale parameter λ is dynamically adjusted according to the length l of the track:

[0028]

[0029] Step 33, missing position prediction: calculate the missing frame position P *

[0030] P * =K(F * , F)(K(F, F)+σ 2 I) -1 P#(4-9)

[0031] Wherein K(,) is the covariance matrix based on k(,), and P is the known track position set.

[0032] Due to the adoption of the above technical scheme, the present application has the following beneficial effects:

[0033] 1. The open vocabulary multi-target tracking method for foggy traffic scenes of the present application proposes an open vocabulary detector OVDetect, OVDetect constructs a RepVL-PAN network to fuse image features and text embeddings, and realizes unknown class recognition through cross-modal alignment; the Prompt-Then-Detect paradigm is adopted: the prompt word embedding is pre-calculated offline, and is directly called during reasoning.

[0034] 2. DIoU-NMS is used instead of traditional IoU, the center point distance constraint is introduced to solve the false suppression caused by the inclusion relationship of the target box, and the missed detection and false detection are effectively reduced.

[0035] 3. The adaptive Gaussian smoothing interpolation (GSI) is introduced, the motion trajectory is modeled based on Gaussian process regression, the smoothing degree is adaptively adjusted by using a dynamic scale parameter, and the broken track is repaired.​ BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a flow chart of the tracking method of the present invention.

[0037] Figure 2 1 is a framework diagram of the open vocabulary detector of the present invention.

[0038] Figure 3 Schematic diagram of DIOU-NMS of the present invention.

[0039] Figure 4 Schematic diagram of Gaussian smoothing interpolation of the present invention. DETAILED DESCRIPTION

[0040] The specific implementation of the invention is further described below with reference to the accompanying drawings.

[0041] Example 1

[0042] The open vocabulary multi-target tracking method for foggy traffic scenarios of the present invention solves the problems of fuzzy target features, unknown categories and broken trajectories in foggy weather, and provides an efficient and robust tracking solution for intelligent transportation and autonomous driving. The method integrates open vocabulary detection and dynamic trajectory optimization to significantly improve tracking performance in complex weather conditions. First, OVDetect based on vision-language joint modeling flexibly identifies targets of unknown categories through RepVL-PAN and region-text comparative learning, breaking through the dependence on fixed categories and coping with foggy feature degradation. Secondly, a DIoU-NMS module based on center distance weighting is proposed to optimize the NMS algorithm, reduce false detections and missed detections, and improve the accuracy of target box screening. Finally, the GSI method based on Gaussian process regression is introduced to capture nonlinear motion patterns, dynamically repair missing trajectories, and enhance trajectory continuity and stability.

[0043] like Figures 1-4 As shown in FIG, the parsing process of the open vocabulary multi-target tracking method for foggy traffic scenes of the present invention includes the following specific steps:

[0044] (1) The input image is processed through an open word detector OVDetect based on joint vision-language modeling to generate object bounding boxes and cross-modal semantic embedding features. OVDetect is an open word detector. OVDetect builds a RepVL-PAN network to fuse image features and text embeddings, and achieves unknown category recognition through cross-modal alignment. It adopts the Prompt-Then-Detect paradigm: prompt word embeddings are pre-computed offline and directly called during inference.

[0045] (2) Redundancy suppression of overlapping target boxes by DIoU-NMS module. DIoU-NMS is used instead of traditional IoU, and the center point distance constraint is introduced to solve the mis-suppression caused by the inclusion relationship of target boxes, effectively reducing the missed detection and false detection.

[0046] (3) Repairing broken trajectories based on Gaussian smoothing interpolation GSI. Introduce adaptive Gaussian smoothing interpolation (GSI), model the motion trajectory based on Gaussian process regression, and use dynamic scale parameter to adaptively adjust the smoothing degree to repair the broken trajectory.

[0047] For step (1), specifically includes:

[0048] (1) Input preprocessing: Obtain the input of foggy image, and optimize the image quality by light dehazing filter module.

[0049] (2) Visual-linguistic feature fusion: RepVL-PAN network performs multi-scale fusion, inputting multi-scale image features X l (l∈{1,2,3} corresponds to the downsampling rate 8 / 16 / 32) and text embedding W j (j is the category word index).

[0050] Text-guided feature enhancement:

[0051]

[0052] Where δ(·) is the Sigmoid function, W j Pre-generated by CLIP text encoder.

[0053] Image pooling attention updates text embedding:

[0054]

[0055] Where is the image feature after global max pooling.

[0056] Reparameterized visual-linguistic path aggregation network (RepVL-PAN) performs multi-scale feature fusion through text-guided CSPLayer.

[0057] (3) Prompt-Then-Detect inference: Predefine an open vocabulary set (such as ["car", "truck", "pedestrian", "unknown_vehicle"]), and generate text embedding offline. Embedding of vocabulary into model weights. Offline generation of embedding vectors of semantic prompt words such as "vehicle" and "pedestrian" reduces real-time computation overhead.

[0058] (4) Output detection results: Generate target bounding boxes and corresponding embedding features.

[0059] For step (2), it specifically includes:

[0060] To enhance the accuracy of subsequent clustering and analysis, we have optimized the data in multiple aspects. Here is a detailed overview of each optimization mechanism:

[0061] (1) Center distance constraint mechanism: Quantify the position deviation between boxes by Euclidean distance square d, solve the problem of traditional IoU failure when there is no overlap (IoU=0). Scale normalization mechanism: Normalize the distance with the diagonal length c of the minimum bounding rectangle, eliminate the influence of target scale difference.

[0062] Redundant box screening. Calculate the DIoU value (formula 4-5) of candidate boxes A and B:

[0063]

[0064] Where, ρ 2 (A, B) is the Euclidean distance square between the center points of the boxes, and C is the diagonal length of the minimum bounding rectangle containing the two boxes. The advantage of DIoU over traditional IoU is that it considers the distance between the boxes, even without overlap, it can provide regression direction and reduce false positives.

[0065] Set the suppression threshold to 0.5: If DIoU>0.5 and score_B<score_A, suppress box B.

[0066] (2) Dynamic threshold adjustment. Dynamic threshold suppression: For adjacent boxes that meet DIoU>θ, keep the highest confidence box and set the rest to zero.

[0067] Based on the dark channel value (range [0, 1]), the fog concentration is estimated in real time. When the fog concentration>0.1, it indicates that the visibility has decreased significantly, and the DIoU threshold is adjusted adaptively to 0.45, further reducing the risk of missing detection. Through the comprehensive use of the above mechanisms, it can more effectively process and optimize log data, and provide a more solid data foundation for subsequent clustering and analysis.

[0068] Fog concentration calculation formula:

[0069]

[0070] Where, I c is the minimum value of the image pixel in the RGB channel, H and W are the height and width of the image.

[0071] Adaptive threshold strategy: Dynamic threshold is adjusted by linear interpolation:

[0072] Threshold = 0.5 - 0.05 max(0, FogDensity - 0.1)

[0073] This strategy improves the target detection recall rate by about 8% in high fog concentration scenes, while keeping the false detection rate below 5%.

[0074] For step (3), specifically includes:

[0075] (1) Trajectory missing frame modeling.

[0076] Based on Gaussian process regression, the missing position prediction constructs a Gaussian process model for the trajectory missing frame:

[0077] f (i) ~ GP(0, k(x, x'))#(4-7)

[0078] (2) Dynamic kernel function adjustment.

[0079] The kernel function uses radial basis function (RBF):

[0080]

[0081] (3) Missing position prediction.

[0082] Input known trajectory position Output missing frame t * position:

[0083]

[0084] Where σ = 0.05 is the noise variance, K is the covariance matrix based on RBF kernel.

[0085] In the Gaussian process regression model construction, the target position function is modeled as P t = f (i) (t) + ∈, where ∈ ~ N(0, σ 2 ) represents noise, and the noise variance σ = 0.05 is determined by grid search. To capture the nonlinear characteristics of target motion, the radial basis function (RBF) kernel is selected as the covariance function, and the dynamic scale parameter λ is calculated adaptively according to the trajectory length and motion pattern, ensuring accurate fitting of the model to complex trajectories.

[0086] Dynamic scale parameter adaptive adjustment: according to the trajectory length l (number of consecutive valid frames), dynamically calculate λ:

[0087]

[0088] Boundary conditions and exception handling:

[0089] Maximum number of consecutive missing frames: ≤ 5 frames (otherwise, consider it as a new target).

[0090] Position confidence evaluation: predicted position standard deviation If σ * >10 pixels, marked as low confidence (need appearance feature assisted verification).

[0091] Motion mutation protection: when the adjacent frame displacement >50 pixels, trigger Kalman filter assisted correction.

[0092] As described above, the tracking method of the present application constructs a complete technical link from target perception to continuous tracking by fusing lightweight open vocabulary detection technology and dynamic trajectory optimization strategy, which significantly improves the tracking performance under complex weather conditions. First, the open vocabulary detector (OVDetect) based on visual-linguistic joint modeling generates cross-modal target proposals through the reparameterization visual-linguistic path aggregation network (RepVL-PAN) and region-text contrast learning, which can flexibly identify unknown class targets not labeled in the training set, breaking through the dependence of traditional detectors on fixed class labels, and effectively dealing with the problems of target appearance feature degradation and background noise interference in foggy scenes. Second, to solve the problem of detection frame redundancy or misjudgment caused by target overlap and feature blur in foggy environment, a DIoU-NMS (Distance Intersection over Union Non-Maximum Suppression) module based on center distance weighting is proposed, which optimizes the traditional NMS algorithm by combining the center point distance and overlapping region information of the bounding box, significantly reducing false positives and false negatives, and improving the accuracy of target frame selection, laying a solid foundation for subsequent trajectory association. Finally, to solve the problem of detection loss and trajectory break caused by occlusion or reduced visibility, an adaptive trajectory interpolation method based on Gaussian process regression (Gaussian Smoothed Interpolation, GSI) is introduced, which captures the nonlinear motion pattern of the target through the radial basis function kernel, dynamically repairs the missing trajectory, and enhances the continuity and stability of trajectory prediction.

[0093] Embodiment 2

[0094] Based on the foregoing open vocabulary multi-target tracking method for foggy traffic scenes, an open vocabulary multi-target tracking system for foggy traffic scenes can be formed, and its application examples and feature combinations are described in the foregoing embodiment 1. The following will be briefly described.

[0095] The open vocabulary multi-target tracking system for foggy traffic scenes of the present embodiment includes the following contents:

[0096] The detection module is used to process the input image through the open vocabulary detector based on visual-linguistic joint modeling, and extract the motion features of the detection results.

[0097] The matching module is configured to match the extracted motion features by calculating the center distance weighted DIoU, and update the matching trajectory.

[0098] The repair module is configured to model the motion trajectory based on the Gaussian process regression by the Gaussian smoothing interpolation processing, and adjust the smoothness adaptively by using the dynamic scale parameter to obtain the repaired trajectory.

[0099] The output module is configured to obtain all trajectories at the current time according to the matching trajectory and the repaired trajectory, and then output the matched trajectory at the current time.

[0100] The specific processing procedure of the open vocabulary detector is as follows:

[0101] Step 11, obtain the target image input, and optimize the image quality by the lightweight defogging filter module;

[0102] Step 12, perform visual-linguistic feature fusion one by one, and the RepVL-PAN network performs multi-scale fusion input multi-scale image features and text embedding, text guided feature enhancement, and image pooling attention updates the text embedding;

[0103] Step 13, adopt the Prompt-Then-Detect paradigm, predefine the open vocabulary set, precompute the prompt word embedding offline, and directly call during inference;

[0104] Step 14, output the detection result, and generate the target bounding box and the corresponding embedding feature.

[0105] The specific processing procedure of the matching module is as follows: the first matching of the extracted high-score detection result motion features, and the second matching of the extracted low-score detection result motion features.

[0106] In the matching module, the DIoU calculation processing procedure is as follows:

[0107] Step 201, quantize the position deviation between the boxes by the Euclidean distance square d;

[0108] Step 202, normalize the distance by the minimum circumscribed rectangle diagonal length c;

[0109] Step 303, for the adjacent boxes satisfying DIoU>θ, keep the highest confidence box and set the rest to zero.

[0110] The Gaussian smoothing interpolation processing procedure of the repair module is as follows:

[0111] Step 31, trajectory missing frame modeling: model the missing frame position P t Construct a Gaussian process regression model:

[0112] P t= f (i) (t)+∈#(4-6)

[0113] where ∈ ~ N(0, σ 2 ) is Gaussian noise, f (i) (t) follows a zero-mean Gaussian process:

[0114] f (i) ~ GP(0, k(x, x'))#(4-7);

[0115] Step 32, dynamic kernel adjustment: radial basis kernel function

[0116]

[0117] where the scale parameter λ is dynamically adjusted according to the length of the trajectory l:

[0118]

[0119] Step 33, missing position prediction: for missing frame position P * calculation:

[0120] P * = K(F * , F)(K(F, F) + σ 2 I) -1 P#(4-9)

[0121] where K(,) is the covariance matrix based on k(,), and P is the set of known trajectory positions.

[0122] As described above, the tracking system of the present application based on the visual-linguistic joint modeling open vocabulary detector (OVDetect) generates cross-modal target proposals through the reparameterized visual-linguistic path aggregation network (RepVL-PAN) and region-text contrast learning, which can flexibly identify unknown class targets not labeled in the training set, break through the dependence of traditional detectors on fixed class labels, and effectively deal with the problems of target appearance feature degradation and background noise interference in foggy scenes. By combining the center point distance and overlapping region information of the bounding box, the traditional NMS algorithm is optimized, significantly reducing false positives and missed detections, and improving the accuracy of target frame selection, laying a solid foundation for subsequent trajectory association. The adaptive trajectory interpolation method based on Gaussian process regression (Gaussian Smoothed Interpolation, GSI) is introduced, which captures the nonlinear motion pattern of the target through the radial basis function kernel, dynamically repairs the missing trajectory, and enhances the continuity and stability of trajectory prediction.

[0123] It should be noted that the above-mentioned examples of the embodiments can be combined with each other as needed, and the multiple examples adopt a set of combined technical features, and the detailed description will not be repeated here.

[0124] It should be noted that the orientations or positional relationships indicated by terms such as "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. are based on the orientations or positional relationships shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation.

[0125] The above description is a detailed description and example of the preferred embodiments of the present application, but these descriptions are not intended to limit the scope of the patent protection claimed by the present application, and any equivalent changes or modified changes completed under the technical guidance of the present application should belong to the scope of the patent protection claimed by the present application.

Claims

1. An open vocabulary multi-target tracking method for foggy traffic scenes, characterized by: Includes the following: Step S1: Process the input image through an open vocabulary detector based on joint vision-language modeling to extract motion features of the detection results; Step S2: Match the extracted motion features by calculating the weighted DIoU based on the center distance, and update the matching trajectory; Step S3: Gaussian smoothing interpolation is used to model the motion trajectory based on Gaussian process regression, and the smoothness is adaptively adjusted using dynamic scale parameters to obtain a repaired trajectory; Step S4: Obtain all trajectories at the current moment based on the matching trajectory and the repaired trajectory, and then output the matching trajectory at the current moment.

2. The open vocabulary multi-target tracking method for foggy traffic scenes according to claim 1 is characterized in that: The specific processing flow of the open vocabulary detector is as follows: Step 11: Obtain the target image input and optimize the image quality through the lightweight dehazing filter model; Step 12: Perform visual-linguistic feature fusion one by one. The RepVL-PAN network performs multi-scale fusion of input multi-scale image features and text embedding, text-guided feature enhancement, and image pooling attention to update text embedding. Step 13: Use the Prompt-Then-Detect paradigm, predefine an open vocabulary set, precompute prompt word embeddings offline, and directly call them during inference; Step 14: Output the detection results and generate the target bounding box and corresponding embedded features.

3. The open vocabulary multi-target tracking method for foggy traffic scenes according to claim 1 is characterized in that: The specific matching process of step S2 is as follows: performing a first matching on the motion features of the extracted high-score detection results, and performing a second matching on the motion features of the extracted low-score detection results.

4. The open vocabulary multi-target tracking method for foggy traffic scenes according to claim 1 is characterized in that: In step S2, the DIoU calculation process is as follows: Step 201: quantify the position deviation between frames using the square of the Euclidean distance d; Step 202: normalize the distance using the minimum circumscribed rectangle diagonal length c; Step 303: For adjacent boxes that satisfy DIoU>θ, retain the one with the highest confidence and set the rest to zero.

5. The open vocabulary multi-target tracking method for foggy traffic scenarios according to claim 1, characterized in that: The Gaussian smoothing interpolation processing flow of step S3 is as follows: Step 31: Modeling the missing frame of the trajectory: The missing frame position P of the i-th target t Construct a Gaussian process regression model: P t =f (i) (t)+∈#(4-6) Among them,∈~N(0,σ 2 ) is Gaussian noise, f (i) (t) obeys a zero-mean Gaussian process: f (i) ~GP(0,k(x,x′))#(4-7); Step 32: Dynamic kernel function adjustment: using radial basis kernel function Among them, the scale parameter λ is dynamically adjusted according to the trajectory length l: Step 33: Missing position prediction: predict the missing frame position P * calculate: P * =K(F * ,F)(K(F,F)+σ 2 I) -1 P#(4-9) Where K(,) is the covariance matrix based on k(,) and P is the set of known trajectory positions.

6. An open vocabulary multi-target tracking system for foggy traffic scenarios, characterized by: Includes the following: Detection module: It processes the input image through an open vocabulary detector based on joint vision-language modeling and extracts motion features of the detection results; Matching module: used to match the extracted motion features and update the matching trajectory through center distance weighted DIoU measurement; Repair module: used to obtain the repair trajectory through Gaussian smoothing interpolation processing, based on Gaussian process regression modeling of motion trajectory, and adaptive adjustment of smoothness using dynamic scale parameters; Output module: used to obtain all trajectories at the current moment based on the matching trajectory and repair trajectory, and then output the matching trajectory at the current moment.

7. The open vocabulary multi-target tracking system for foggy traffic scenarios according to claim 6, characterized in that: The specific processing flow of the open vocabulary detector is as follows: Step 11: Obtain the target image input and optimize the image quality through the lightweight dehazing filter model; Step 12: Perform visual-linguistic feature fusion one by one. The RepVL-PAN network performs multi-scale fusion of input multi-scale image features and text embedding, text-guided feature enhancement, and image pooling attention to update text embedding. Step 13: Use the Prompt-Then-Detect paradigm, predefine an open vocabulary set, precompute prompt word embeddings offline, and directly call them during inference; Step 14: Output the detection results and generate the target bounding box and corresponding embedded features.

8. The open vocabulary multi-target tracking system for foggy traffic scenarios according to claim 6, characterized in that: The specific matching processing flow of the matching module is as follows: performing a first matching on the motion features of the extracted high-score detection results, and performing a second matching on the motion features of the extracted low-score detection results.

9. The open vocabulary multi-target tracking system for foggy traffic scenarios according to claim 6, characterized in that: In the matching module, the DIoU measurement process is as follows: Step 201: quantify the position deviation between frames using the square of the Euclidean distance d; Step 202: normalize the distance using the minimum circumscribed rectangle diagonal length c; Step 303: For adjacent boxes that satisfy DIoU>θ, retain the one with the highest confidence and set the rest to zero.

10. The open vocabulary multi-target tracking system for foggy traffic scenarios according to claim 1, characterized in that: The Gaussian smoothing interpolation processing flow of the repair module is as follows: Step 31: Modeling the missing frame of the trajectory: The missing frame position P of the i-th target t Construct a Gaussian process regression model: P t =f (i) (t)+∈#(4-6) Among them,∈~N(0,σ 2 ) is Gaussian noise, f (i) (t) obeys a zero-mean Gaussian process: f (i) ~GP(0,k(x,x′))#(4-7); Step 32: Dynamic kernel function adjustment: using radial basis kernel function Among them, the scale parameter λ is dynamically adjusted according to the trajectory length l: Step 33, missing position prediction: missing frame position P * calculate: P * =K(F * ,F)(K(F,F)+σ 2 I) -1 P#(4-9) Where K(,) is the covariance matrix based on k(,) and P is the set of known trajectory positions.

Images