Research On Method Of Vehicle Object Tracking Based On Correlation Filter

The increasing of car ownership makes traffic congestion and traffic safety more serious.With the development of visual tracking,the application of target tracking technology to road traffic scenes can effectively solve traffic congestion and traffic safety problems.As one of the key technologies of intelligent transportation system and automatic driving system,vehicle target tracking has important research significance and application value.In view of the change of target scale and the complexity and changeability of road traffic scene in the tracking process,this paper designs a target tracking algorithm suitable for vehicles based on the kernel correlation filtering tracking algorithm,which is improved from two aspects of scale estimation and feature extraction.The main work of this paper is as follows:Firstly,a correlation filtering tracking algorithm based on adaptive aspect ratio is proposed.Aiming at the scale change caused by the rapid movement of the target in the vehicle tracking scene,the forward differential prediction algorithm was used to predict the target scale of the current frame.Based on the prediction scale,the scale pool was constructed to perform the scale estimation of the adaptive aspect ratio.Then,the forward differential prediction algorithm is adjusted by the target scale and the prediction scale to ensure the accuracy of the scale prediction.At the same time,the average peak energy and the maximum response value are used to realize the adaptive update of the model,and the learning rate of the model is dynamically adjusted according to the target velocity,so as to reduce the possibility of model drift due to background interference in the tracking process.Experimental results on OTB100 and Car Data data sets show that this algorithm breaks the restriction of fixed aspect ratio in the tracking process.Compared with the five classical tracker,the tracking accuracy and success rate of the algorithm have been improved to a certain extent,and it can better deal with the rapid movement and scale change of the target in the process of vehicle tracking.Then,from the perspective of features,a multi-feature fusion correlation filter tracking algorithm combining context awareness is proposed to build a more robust appearance model to improve the discriminant power of the tracker in vehicle tracking scenes.In this algorithm,the traditional manual features FHOG and CN are fused,and the deep convolution features extracted from VGGNET-19 model are trained with the correlation filter respectively to detect the target.The adaptive fusion of these two features is carried out at the decision layer to determine the target position.Through the complementary advantages of the two,the characterization ability of target features is improved.At the same time,contextual information is introduced into the filter training stage to make full use of background information to improve the discriminant ability of relevant filters.Experimental results on OTB100 and Car Data data sets show that compared with the five classical trackers,the tracking accuracy and success rate of this algorithm are improved,and it can better adapt to complex and changeable traffic scenes.