YOLO-based specific target tracking method

Abstract

The invention provides a YOLO-based specific target tracking method, and belongs to the technical field of target tracking. The method comprises the following steps: firstly, setting a to-be-tracked specific target, shooting a video in real time by using a camera, sending the video to a computer, and putting an image of the specific target as a template into a template folder; Taking the first frame of image with the specific target in the video as a first frame of image, and carrying out target detection on each frame of image by using a YOLO detection algorithm to obtain a detection result corresponding to the type of the object where the specific target is located in each frame of image; And correcting the detection result by using a Deepsort algorithm, and then tracking a specific target by using a Surf algorithm on the corrected result. According to the method, continuous tracking of a single specific target can be realized, the speed is high, the real-time performance is good, and the method is well suitable for environment changes.

Description

technical field [0001] The invention relates to the technical field of target tracking, in particular to a specific target tracking method based on YOLO. Background technique [0002] At present, with the development of computer vision technology, vision-based moving target detection and tracking has become a current research hotspot, and has broad application prospects in video surveillance, virtual reality, human-computer interaction, planetary detection, and precision guidance. [0003] YOLO is an object detection algorithm based on deep learning neural network architecture. The algorithm divides the input image into 7 by 7 grids, and presets 5 default boxes centered on each grid. The output of the algorithm is in The offset based on the 5 default boxes is predicted at each grid, and the corresponding category is predicted at the same time. The 5 preset default boxes are 5 representative boxes obtained by clustering on a large number of object detection data sets, thus e...

AI Technical Summary

Problems solved by technology

[0004] Most of the existing YOLO-based target tracking methods cannot track a specific single target, but can only track the same type of object, and cannot distinguish between different individuals in the same type of object, and the existing YOLO-based target Tracking methods also do not account for occluded objects

For example, in 2018, Zhu Chenyang from Nanjing University of Science and Technology proposed a YOLO3-based automatic face tracking camera robot system research. They designed an automatic tracking camera robot and proposed a deep convolutional neural network based on YOLO3. The method of cyclic fast detection and tracking, but this method can only be applied in the case of a single face, and cannot distinguish different faces; in 2018, Cai Chengtao et al. of Harbin Engineering University proposed a panoramic multi-target real-time detection based on the improved YOLO algorithm Research, they proposed a real-time detection method for panoramic targets based on the improved YOLO algorithm, but this method still cannot distinguish between different targets, and when the target is occluded, the tracking accuracy is reduced by 6 percentage points

Images