Research On Deep Learning Based Vehicle Tracking

With the rapid economic growth,the number of vehicles have appeared in an explosive growth.This has provided convenient and brought many traffic problems at the same time.The deep learning and big data-related technologies develop rapidly,this problem of traffic system could be solved.How to identify and track specific goals in complex scenes is crucial in the development of intelligent transportation in the future.How to adaptively track real-time and accurate in vehicle tracking for various real scenarios has always been a difficult area for academic research.This article is based on actual traffic video scenes,based on deep learning algorithms including SSD(Single Shot Multi-Box Detector)and CNN(Convolutional Neural Networks),based on a variety of actual scenes designed and realized with both real-time and wth the accuracy of vehicle tracking scheme,a multi-module vehicle tracking scheme based on deep learning was proposed,and the entire program was deployed on the big data platform.The main research contents are as follows:(a)Firstly,we study and analyze the traditional algorithms and deep learning methods in processing target detection,image feature extraction,and applying the method to the sub-modules of the vehicle tracking program.Because of deep learning has the characteristics of strong adaptability,large model capacity,etc.,it can perform adaptive learning for different scenarios and target features and establish suitable features.These features will reduce manual interventions,while improving the accuracy of detection and identification,providing a solid foundation for intelligent transportation systems.(b)Secondly,the overall scheme of vehicle tracking is proposed,including the detection of moving vehicles based on SSD,the extraction of vehicle features based on CNN and vehicle model recognition,and the use of Euclidean distance combined with vehicle tracking of spatio-temporal information.Vehicle detection and vehicle model recognition all use deep learning methods to analyze their models,and for real-time data sets,multiple trainings and improvements to their model parameters make the accuracy of detection and recognition greatly improved.(c)The Last,deployed the scheme on the video big data processing platform that combines Hadoop and Storm with high speed and high data processing capability to ensure real-time vehicle tracking.Different from the traditional method of tracking image feature information,this solution is combined with multiple deep learning implementations.The experimental results show that while ensuring real-time performance,the new scheme has better tracking effect for more vehicles in the video,which verifies the effectiveness of the scheme.In order to verify the feasibility and portability of this scheme,the vehicle tracking scheme was transplanted to the embedded development board.The experimental results show that this scheme has strong feasibility and portability.