Multi-target tracking method based on association detection and coding network

Abstract

The invention provides a multi-target tracking method based on correlation detection and a coding network, and belongs to the field of image recognition. The method comprises: taking a target detection frame meeting a preset condition in target detection frames of each frame of image of a video stream as a key target detection frame, and determining an identity feature vector of the key target detection frame; calculating the distance between the identity feature vector of the key target detection frame of each frame of image and the identity feature vector of the key target detection frame of the previous frame of image in the Euclidean space to form a first incidence matrix so as to determine the incidence relation between the key target detection frame of each frame of image and the key target detection frame of the previous frame of image; and correspondingly associating the key target detection frame of each frame of image with the key target detection frame of the previous frame of image to form a motion track of each key target detection frame of each frame of image in the video stream. The problem that a multi-target tracking method cannot be reasonably trained together during joint training detection and identity feature coding is solved.

Description

technical field [0001] The invention relates to the field of image recognition, in particular to a multi-target tracking method based on an association detection and coding network. Background technique [0002] Multi-object tracking methods are widely used in the fields of autonomous driving, behavior recognition, and mobile robots, mainly by dividing the tracking into two steps, namely, the detection step (locating the target independently in each frame) and the data association step (the neural network learns the identity discriminative features and frame-by-frame associated bounding boxes to form trajectories) to simplify the task. Existing multi-object tracking methods reduce the complexity of overall object tracking and turn multi-object tracking into a data association problem. [0003] Detection in multi-object tracking methods inherently suffers from missed and false detections, and in crowded environments where objects are occluded and interact with each other, so...

Images