Research On Vehicle Tracking Algorithm Based On Deep Learning

In recent years,the number of motor vehicles has grown rapidly,and existing transportation resources have not been able to meet daily traffic demands.The intelligence of traffic monitoring can relieve a certain degree of traffic pressure.The vehicle tracking system provides the key content for intelligent traffic monitoring and is the basis for obtaining a variety of traffic parameters.As the actual traffic environment is more complex,such as changes in light illumination,scale changes,obstructions,and cluttered backgrounds,it is very difficult to obtain a precise,stable,robust vehicle tracking method.The rapid development of deep learning has provided new ideas for the research of intelligent traffic monitoring systems.Applying deep learning to vehicle tracking can effectively improve the performance of vehicle tracking algorithms.The video tracking method based on video images studied in this paper belongs to single-target tracking and provides traffic monitoring systems with accurate and robust tracking of specific vehicles.Firstly,the background of the vehicle tracking system and the research results in the field of vehicle tracking are analysed.Then the current mainstream tracking algorithm framework is studied,and the modules under the framework are analysed.Then the kernelized correlation filter-tracking algorithm is analysed in detail.It is a good tracking algorithm.The feature extraction module in the tracking algorithm has a very important role.The convolutional featurest is extracted by the CNN.Compared with the HOG features used in the kernelized correlation filter algorithm,the convolution features can describe the more essential features of the target.Firstly,the principle and structure of convolutional neural network are studied.Based on the VGG-Net network model trained on ImageNet dataset,the features extracted from each convolution layer are analysed hierarchically,and compare the feature maps of different layers through experiments.Select the conv3-3 layer to retain better underlying detail features and conv5-3 layers of high-level semantic features.Then use the principal component analysis method to adaptively reduce the the low-level features,thereby reducing the amount of calculations.Based on the kernelized correlation filter-tracking algorithm,select two-layer convolution features to train filter.The coarse-to-fine positioning method is used to synthesize the output of the two filters to achieve accurate positioning of the tracking vehicle.The kernelized correlation filter-tracking algorithm does not achieve the target scaling adjustment.This paper uses a scale filter to recalculate the size of the tracking veh to adapt to the changes in the scale of the target vehicle icle.The frequency of model updating of thethis algorithm is too frequent,which easily leads to model pollution.By performing the peak-to-peak ratio calculation of the tracking position,the tracking result with the PSR value not lower than the threshold value is selected for model updating.Select the vehicle video in the VTB-100 data set for experiments.Experiments show that the vehicle-tracking algorithm proposed in this paper can better track the vehicle and is robust to motion blur?occlusion and scale changes.