Research On CNN Particle Filtering Algorithm For Vehicle Tracking

With the rapid development of China's economy,the number of private cars has exploded.People have also encountered many transportation problems while being facilitated.With the advent of deep learning and big data related technologies,related transportation problems are expected to be resolved.In the future development of intelligent transportation,how to accurately find the target vehicle in massive videos and in the case of more complicated road traffic scenarios is very important.Tracking vehicles for various actual scenarios,real-time and accuracy have always been difficult research points.This paper proposes a particle filter tracking algorithm based on CNN feature extraction to solve the problems of robustness and real-time performance of tracking and improve the accuracy of tracking.Firstly,the convolutional neural network model was studied,and CNN was used to extract features,and its feature extraction method was improved.Aiming at the shortcomings of traditional pooling algorithm feature extraction,which is not flexible enough,a chaotic variable-scale firefly pooling algorithm was proposed.The variable-scale firefly algorithm introduces the pooling algorithm,and continuously optimizes the pooling parameters until it converges,making it have certain flexibility.The improved algorithm can overcome the inherent shortcomings of the original pooling algorithm.Secondly,Faster-RCNN detection method is used for feature extraction and target detection.The improved CNN algorithm is used to obtain the convolution feature map and add it to the RPN network.It is used to find the suggested area that may contain the target,the target score is used to filter the suggested area box,and then the target types in the area are classified.Discriminate and return to the preselection box that belongs to a certain feature,and further adjust its position.Finally,the tracking algorithm uses Mean Shift and weight-optimized particle filtering method,and uses the clustering effect of Mean Shift to converge the particle samples to the real position closest to the target.During the resampling process of the particle filtering algorithm,particles with large weights will be copied,andcorrespondingly discarding particles with small weights will reduce the diversity of particles.Therefore,a weight optimization algorithm is used to optimize the particle weights to improve the diversity of particles.Then use the target detection and particle filter algorithm to predict the target trajectory.The Hungarian algorithm performs the target correlation between the front and back frames to obtain the best tracking trajectory and complete the tracking experiment.It is verified on public data sets that the algorithm in this paper achieves good accuracy and real-time performance,effectively reduces the amount of calculations,and can well adapt to lighting changes,complex backgrounds,and occlusions.