Research On Methods For Moving Vehicle Tracking In Mobile Video

The effective information extracted from videos has always been an important basis for humans to make decisions.How to use tools to automatically extract effective information from videos has always been pursued by humans.Target tracking technology has always been one of the important research directions in the field of machine vision research,and it has a large number of applications in multiple real-life scenarios such as automatic driving,flight control,video surveillance,and drone reconnaissance.The main task of video target tracking is to manually mark the target to be tracked in the first frame of a given video sequence,and then accurately and quickly estimate the target’s position without knowing the target’s position in the subsequent frames.However,in actual application scenarios,there are many interference factors that affect the tracking performance of the tracking algorithm,such as target occlusion,rapid target movement,camera shake,similar background targets,motion blur,and out of view.Therefore,designing a target tracking algorithm with high accuracy,good real-time performance and good robustness is still a challenging subject.Based on the Mean Shift algorithm,compressed tracking algorithm,and siamese network,this paper will conduct research from key parameter adaptation,feature representation,template update,loss function,and model fusion to improve the performance of the tracking algorithm.The main work of this paper is as follows:1.A vehicle tracking algorithm based on the mean shift of chaotic particle swarms is proposed.In view of the difficulty in selecting the bandwidth of the kernel function in the mean shift algorithm,the particle swarm algorithm is used to adapt the kernel function bandwidth,and the particle swarm algorithm is easily limited to the local optimal problem.On the basis of it,chaotic variables are integrated,and the characteristics of the chaotic variables can make the particles The group algorithm jumps out of the local optimal state,so as to achieve the goal of global optimal.2.An improved compressed vehicle tracking algorithm based on sparse representation is proposed.Aiming at the shortcomings of the classic target compression and tracking algorithm,it has been improved.In the feature selection,the principal component analysis method is proposed to select high-discrimination features.Then a template update strategy based on Bhattacharyya distance is proposed,and finally a coarse and precise search strategy is designed to adapt the size of the target window.3.A vehicle tracking algorithm based on the improved Siamese FC network is proposed.Firstly,the fusion of the features of the multilayer network is used to obtain the features with good robustness.Aiming at the imbalance problem of positive and negative samples and the problem of difficult samples,the loss function is improved so that the loss function pays more attention to small-class samples and difficult samples.4.A vehicle tracking algorithm based on the Siamese FC network based on the mean shift algorithm is proposed.The tracking confidence method is adopted to update the target vehicle area template to adapt to changes in the shape of the target vehicle and changes in background information.Aiming at the video frame of the target vehicle drifting phenomenon tracked by the improved Siamese FC network,the position of the target vehicle is optimized by introducing the mean shift algorithm.