Multiple Object Tracking Algorithm Based On Machine Learning

With the development of science and technology,multi-target tracking has become a more popular research topic and an important research direction in the field of machine vision,which is widely used in military and civilian fields.The purpose of multi-target tracking is to continuously track multiple target objects,during which the label of the same target is maintained unchanged,and the state of each target in the future frame is predicted.This paper mainly focuses on the problem of poor tracking effect of multi-target tracking algorithm when the target is occluded,and improves on the existing algorithm to obtain excellent tracking effect in the case of occlusion.The main research contents of this paper are as follows:1)This paper improves on the Mask R-CNN algorithm and proposes a multi-scale detection algorithm for small target deformation.First,the backbone network is changed from Res Net-101 network to Res Ne Xt-101 network,which improves the network detection accuracy without increasing the complexity;Secondly,a top-down reverse connection is added to the traditional FPN to improve the accuracy of small target detection;Finally,the offset is added to the Ro I Align in the Mask R-CNN algorithm to improve the target detection accuracy in the presence of occlusion.The detection algorithm is experimentally validated on the pedestrian dataset MOT15 and the self-photographed data set.The experimental results demonstrate that the improved detection algorithm in this paper improves the accuracy of target detection in the presence of small targets and occlusion.2)In this paper,a trajectory correction multi-target tracking algorithm is proposed based on the SORT algorithm.First,the Hungarian algorithm,which only uses Io U as the benchmark,is changed to a data association algorithm that uses Io U and color histograms to improve the accuracy of matching;Secondly,in view of the SORT algorithm’s inability to deal with the problem of target occlusion,a trajectory correction algorithm is proposed,which reduces the problem of label switching before and after occlusion.The tracking algorithm is experimentally validated on the pedestrian dataset MOT16 and the self-photographing dataset.The experimental results demonstrate that the improved multi-target tracking algorithm in this paper can improve the tracking performance during occlusion with almost no sacrifice of speed.