Research And Implementation Of Vehicle Detecion Algorithm Based On Deep Learning

With the rapid development of economy,urban population and automobile ownership are increasing dramatically,which leads to heavy traffic and makes traffic jam more and more serious.In order to reduce the traffic pressure,the most critical point is to count accurate statistics of the traffic flow to achieve a reasonable traffic diversion,and the vehicle automatic detection is a crucial step to achieve this goal.Besides,in recent years,a substantial increase of road surveillance camera also provides a hardware condition for vehicle detection.Therefore,it is of great importance to carry out further research on real-time video automatic detection of vehicle and improve its detection accuracy.The traditional method of video surveillance image has a great disadvantage,interference of background and light,inclement weather like heavy snow and rain,overlapping vehicle monitoring in video image,will greatly affect the accuracy of vehicle detection.What's more,the traditional method can't meet the requirement of real time because of its time consuming.As one of the most popular technology,deep learning technology has shown its amazing ability in many fields.Using convolutional neural network to detect the image is a great application of deep learning in image processing.In this paper,the traditional methods of vehicle detection are summarized,and a new detection algorithm based on the application of deep learning in vehicle detection is proposed for the first time,which aims at the traditional's shortcomings and focuses on solving the problem that the traditional's can't solve.Firstly,the traditional algorithms of classic vehicle detection are studied deeply,and their theories are introduced in details.After comparing their detection results of complex scenarios and analyzing their advantages and disadvantages,this paper comes to a conclusion that traditional algorithms can't meet requirements of complex scene mentioned in the paper.Therefore,a new algorithm of vehicle detection based on deep learning technology is proposed.RPN network and Fast RCNN network are trained separately and alternately to achieve volume layer sharing,so as to realize the detection of the vehicle to though a unified network.Secondly,characteristics of deep learning framework Caffe,software tools and complex steps of framework building are introduced.This paper explains the network model in Caffe framework,including the design of network parameters and training parameters.Videos taken in different weather conditions like sunny,cloudy,rainy,snowy and evening,are converted to 20000 pictures.In order to get the self-made data,vehicles in these pictures are tabbed.Then,these pictures are trained by gradient descent method.So the vehicle detection network is realized.At last,this algorithm and traditional algorithms are tested with sample pictures in six of different scenes.Then this paper analyzes each algorithm from detection accuracy and real-time performance,as well as their advantages and disadvantages.Compared with traditional algorithms of classic vehicle detection,this algorithm has higher accuracy,better real-time performance and stronger anti-interference ability against background,light,bad weather and vehicle overlap.