Multi-Target Pedestrian Stereo Target Detection And Tracking Based On Binocular Camera

Today’s multi-target detection and ranging technologies are often used in a variety of industrial applications,such as part alignment in construction sites and collision detection for pedestrians.Among the traditional detection methods,the most currently used detection and tracking techniques for these applications are laser infrared radar ranging and moving object detection on visual images.However,these methods can be not only susceptible to complex backgrounds or occlusions,but also difficult to popularize due to the high price of the equipment.In order to remedy the shortcomings in the traditional methods,this paper improves the ranging part of the traditional methods and uses Reid reidentification network to solve the problem of multiple target matching in binocular cameras under the premise of using binocular vision instead of infrared radar ranging,and enhances the ability of algorithmic target tracking using distance information and improved neural network.The specific implementation of this paper is based on the improved YOLOv3 and Deep Sort algorithms,and the effect of being able to detect the stereo coordinates of multiple pedestrian targets and plot the motion trajectory in different scenes is achieved using the principle of binocular vision.The main work and research are as follows:(1)Target pairing using Reid feature extraction network.Since the objects are multiple targets,the targets in the left and right images in the binocular camera must be paired before depth information measurement is performed.In this paper,we use the Reid feature extraction network to extract the targets into feature vectors and then perform target matching by similarity.(2)Improve the Kalman filter in Deep Sort algorithm.Adding a dimension of depth information detection improves the original object tracking ability of the algorithm while optimizing the accuracy of depth information through Kalman gain.(3)Improving the target detection network YOLOv3 and the neural network part in the multi-target tracking algorithm Deep Sort by improving the target recognition ability and tracking ability of the algorithm.