CN-116258746-B - Target tracking method, target tracking system, electronic equipment and storage medium

CN116258746BCN 116258746 BCN116258746 BCN 116258746BCN-116258746-B

Abstract

The invention discloses a target tracking method, a target tracking system, electronic equipment and a storage medium, which can prevent the problem of tracking frame deviation caused by long-time tracking by correcting a target by adopting a detector in the target tracking process. In addition, the yolov target detection framework is improved, so that the model is provided with a lightweight 16-layer backbone network, the overall weight of the model is lighter under the condition that the richness of extracted features and the detection accuracy are ensured, real-time target tracking can be performed under the condition that the equipment performance is low and the equipment computing power is low, and the possibility of following an error target in the target tracking process can be reduced based on modeling of target colors. The embodiment of the invention can be widely applied to the technical field of computers.

Inventors

LIANG JUN
ZHOU NUO
LONG JIAHAO
LUO RUI

Assignees

华南师范大学

Dates

Publication Date: 20260512
Application Date: 20230203

Claims (10)

1. A method of tracking a target, comprising: Constructing a detector based on a modified yolov target detection framework, wherein the modified yolov target detection framework has a 16-layer backbone network; constructing a tracker based on modeling of target colors and a CSRT target tracking algorithm; Acquiring a video stream and an initial frame of the video stream; initializing the tracker according to the initial frame; configuring a tracking time condition, wherein the tracking time condition comprises a tracking timer and a searching time limit; According to the video stream and the tracking time condition, carrying out target tracking on a target object through the tracker; in the target tracking process, correcting an image area where the target object is located through the detector; The improvement yolov target detection framework includes an improvement yolov-mobile target detection framework configured with a dynamic learning mechanism; the dynamic learning mechanism is specifically as follows: Pooling output by global averaging Performing a Squeeze operation on the pooling layer of the database; and carrying out learning excitation operation through the full connected layer, leakyReLU activation functions, the full connected layer and the h_sigmoid activation function combination.
2. The method of claim 1, wherein said constructing a detector based on a modified yolov target detection framework comprises: Constructing an improved yolov model by adopting an improved yolov target detection framework, wherein the improved yolov model has a 16-layer backbone network; model training is carried out on the improved yolov model to obtain a target detection model; and constructing a detector by combining a dynamic learning mechanism and the target detection model.
3. A method of object tracking according to claim 1, wherein said configuring said tracker comprises: Intercepting an initial frame of the video stream; Determining a target area where a target object is located; determining a target object contour by the detector based on the initial frame and the target region; initializing the tracker according to the initial frame and the target object outline.
4. The method according to claim 1, wherein the target tracking of the target object by the tracker according to the video stream and the tracking time condition comprises at least one of: When the target object does not exist, storing a target model in an image frame of the target object which finally exists, starting searching time limit timing, and detecting the target of the image frame in the searching time limit by the detector until the target object is detected or the searching time limit is reached; When the target object exists and the tracking timer is over, performing target correction on the image frame through a target correction function of the detector, resetting the tracking timer, and continuing to perform target tracking on the target object through the tracker after resetting the tracking timer; and when the target object exists and the tracking timer is not finished, continuously tracking the target object through the tracker.
5. The method according to claim 4, wherein the detecting, by the detector, the object of the image frame within the search time period until the object is detected or the search time period is reached, comprises: Stopping searching time counting when the target object is detected, and continuing to track the target object through the tracker; when the searching time limit is up, an abnormal alarm is sent out, and the target tracking is stopped.
6. A method of object tracking according to claim 4, characterized in that the method further comprises: And correcting the target model according to the change condition of the target area tracked by the target.
7. The method according to claim 1, wherein the target tracking of the target object by the tracker further comprises: acquiring target area movement information between image frames; And predicting the position of the target area in the next image frame according to the movement information of the target area.
8. A target tracking system, comprising: A first module for constructing a detector based on a modified yolov7 target detection framework, wherein the modified yolov target detection framework has a 16-layer backbone network, and the detector is used for correcting the area of a target object; a second module for constructing a tracker based on modeling of target colors and a CSRT target tracking algorithm; a third module, configured to obtain a video stream and an initial frame of the video stream; a fourth module for initializing the tracker according to the initial frame; The fifth module is used for configuring a tracking time condition, wherein the tracking time condition comprises a tracking timer and a searching time limit; a sixth module, configured to perform target tracking on the target object by using the tracker according to the video stream and the tracking time condition; a seventh module, configured to correct, in a target tracking process, an image area where the target object is located by using the detector; The improvement yolov target detection framework includes an improvement yolov-mobile target detection framework configured with a dynamic learning mechanism; the dynamic learning mechanism is specifically as follows: Performing a squeze operation through a pooling layer of global average pooling output; and carrying out learning excitation operation through the full connected layer, leakyReLU activation functions, the full connected layer and the h_sigmoid activation function combination.
9. An electronic device comprising a processor and a memory; the memory is used for storing programs; The processor executing the program implements the method of any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that the storage medium stores a program that is executed by a processor to implement the method of any one of claims 1 to 7.

Description

Target tracking method, target tracking system, electronic equipment and storage medium Technical Field The present invention relates to the field of computer technologies, and in particular, to a target tracking method, a target tracking system, an electronic device, and a storage medium. Background In recent years, due to the rapid development of computer vision, object tracking, which is one of the cores in the field of computer vision, is also becoming more and more important. However, in current target tracking algorithms, it tends to perform poorly in scale transformation and accumulated offset caused by occluded scenes and long-term tracking of targets. While for these problems, the academy proposes a tracking method based on deep learning. Earlier target tracking algorithms were based primarily on target modeling or tracking target features. The method according to object modeling requires modeling an object appearance model, and finding objects in a sequence of subsequent frames, wherein the representative algorithm is an optical flow method. The method for tracking the target features is to firstly extract the target features by using a feature matching method, and then find the most similar features in the subsequent frames for positioning. A multi-instance learning algorithm (MILs) has been proposed to develop a discriminant model by packing all ambiguous negative and positive samples. Related filtering-based tracking algorithms have been proposed in combination with the communication field, among which the KCF and MOSSE algorithms are known for their high real-time performance, while the CSRT algorithm exhibits high accuracy. At the same time, there are also learners to break down the final task of tracking into tracking, learning and detection (TLD) subtasks, where tracking and detection strengthen each other, thus proposing TLD algorithms. Along with the combination with the deep learning field, the target tracking algorithm has a leap type progress, siamRPN series tracking algorithms, and the tracking accuracy is greatly improved through the design of a twin network. However, in practical applications, although the algorithm can achieve maximum power point tracking, there are respective disadvantages, mainly: (1) The optical flow method and the feature matching method are easy to cause tracking failure under the interference of target shielding, illumination change, motion blurring and the like because the background information is not taken into consideration in the algorithms. Meanwhile, the algorithm execution speed is low, and the requirement of real-time performance cannot be met. (2) The MIL, KCF, TLD, MOSSE algorithm performs poorly in cases where it contains target occlusion and scale transformation and accumulated offset due to long-term tracking, while the accuracy and power consumption of the CSRT algorithm are quite dependent on the framing of the target, with less robustness. (3) The algorithm of pure deep learning represented by SiamRPN has high requirements on running equipment, and occupies too high CPU and memory in low-performance equipment such as a mobile phone, a flat panel or a part of low-price chips, is basically difficult to run, and cannot be well applied to industry. These methods have high performance requirements on computing devices, thus making the actual investment costs high. Meanwhile, for objects such as unmanned planes, robots and the like using lightweight computing equipment, the constraints on equipment performance make it difficult for the unmanned planes, robots and the like to apply a deep learning method to track targets in real time. Disclosure of Invention Therefore, the embodiment of the invention provides a target tracking method, a system, electronic equipment and a storage medium, which have high accuracy, strong real-time performance and low equipment performance requirements. The embodiment of the invention provides a target tracking method, which comprises the steps of constructing a detector based on an improved yolov7 target detection frame, constructing a tracker based on modeling of target colors and a CSRT target tracking algorithm, acquiring a video stream and an initial frame of the video stream, initializing the tracker according to the initial frame, configuring tracking time conditions, wherein the tracking time conditions comprise a tracking timer and a searching time limit, tracking a target object through the tracker according to the video stream and the tracking time conditions, and correcting an image area where the target object is located through the detector in a target tracking process. Optionally, the construction of the detector based on the improved yolov target detection framework comprises the steps of constructing an improved yolov model by adopting an improved yolov target detection framework, wherein the improved yolov model is provided with a 16-layer backbone network, performing model training on the improved yo