Research On Object Tracking Methods Based On Lightweight Deep Neural Networks

Object tracking is a basic task in the field of computer vision,and its application scenarios are very extensive,including automatic driving,military security,human-computer interaction and other aspects.The essence of target tracking is to predict the trajectory of the moving target in the video,and the true position of the tracked target is only specified when the target initially appears.The difficulty of tracking problems is that the target will change during tracking,such as deformation,occlusion.In the context of the rapid development of convolutional neural networks,more and more tracking algorithms have been proposed,including many tracking methods with superior performance.Although most tracking methods have good performance,their performance improvement relies on heavy parameter network models,which makes it difficult for tracking algorithms to track in real time.Conversely,some relatively lightweight tracking methods are difficult to perform well in a series of complex backgrounds.For the existence of the above tracing problems,the thsis aims to build a lightweight tracing algorithm.The main research work is divided into the following three parts:1.Build a lightweight backbone network,obtain the optimal network structure through ablation experiments,and the initialization weights are obtained by pre-training on the COCO dataset.Based on the backbone network,a lightweight twin network is constructed according to the correlation filtering principle,and the influence of different region filling methods on tracking performance is explored by comparative experiments.The feature of the target area and the search area are used to perform correlation operations,and the maximum value of the response is mapped back to the original map to obtain the current position of the tracking target,so as to realize the target tracking.2.The reasons for the poor monetization of the twin network when the target is deformed during the tracking process are analyzed.Under the condition of ensuring the lightweight of the network,multi-scale features are added for tracking.Features at different scales contain semantic features at different levels,and feature pyramids are used to integrate these semantic features to strengthen the extraction of target features.The extracted target features are fed into the regional recommendation network,and the regression obtains the target position.Finally,the effects of adding feature pyramids and regional recommendation networks on tracking performance are verified by experiments.3.For target tracking,the tracking object briefly disappears and then reappears in the picture,causing the tracking failure.Add a dual-stream network to the lightweight backbone network to model the tracked target in time and space.After the target disappears,the target feature information and spatial location information are retained for a short time,and the target is not deleted until the target does not appear for a period of time.Experiments show that the dual-stream network can improve the tracking performance when the target disappears briefly.