A Multi-Object Tracking and Segmentation Method Using Short-Range Association and Long-Range Pruning

Abstract

A multi-object tracking and segmentation method using short-range association and long-range pruning, involving computer vision. Train the convolutional neural network for segmentation and tracking, input the video picture into the trained network, and obtain the segmentation position and instance representation vector corresponding to each target instance in the video image; use the Euclidean distance to measure the space of representation vectors between different instances The distance and the instance mask center distance are converted into vector similarity score and mask center similarity score; the mask similarity score and edge box similarity are obtained by using the instance mask and the propagation score of the edge box between adjacent frames Score; use four similarity scores and the Hungarian algorithm to get the motion trajectory of the target instance in the video; in the target instance trajectory, use the target instance confidence score of the previous frame to adjust the target instance confidence score of the current frame, and clear the low instance confidence score trajectories, and obtain long-range motion trajectories with high confidence scores. It has high precision and robustness.

Description

Technical field [0001] The present invention relates to the field of computer vision technology, in particular to a multi-objective tracking and segmentation method using short-range correlation and long-range pruning. Background [0002] In the field of computer vision, the use of convolutional neural networks for multi-objective tracking and segmentation has been a great success. The recent multi-object tracking method is mainly to apply the paradigm of first detection and then tracking, and to link the detection target together to form a motion trajectory by applying a data correlation algorithm. The performance of data correlation is highly dependent on the similarity measurement method, but the similarity measurement is not stable due to frequent occlusion between targets. Chinese Patent CN201910176113.X discloses an online multi-target tracking method based on trajectory metric learning, for solving the technical problems of poor practicality of existing online multi-target...

AI Technical Summary

Problems solved by technology

While the performance of data association is highly dependent on the similarity measurement method, but due to frequent occlusions between objects, the similarity measurement is not stable

Chinese patent CN201910176113.X discloses an online multi-target tracking method based on trajectory metric learning, which is used to solve the technical problem of poor practicability of existing online multi-target tracking methods

However, it is difficult to obtain long-range trajectories with high confidence due to the lack of spatio-temporal information in adjacent frames (i.e., local trajectory information, which is beneficial for constructing short-range trajectories)

Images