Real-time Vehicle Type Matching Based On Deep Learning

With the rapid development of intelligent transportation systems and the massive deployment of transportation surveillance cameras,traffic videos increase in a rapid way.Real-time vehicle type matching can be very important in emergency situations for security reasons,such as traffic accidents identification from traffic videos.However,it is difficult for traditional video processing methods to real-time handle video data and feedback results.Traditional traffic video processing methods provide the service of video storage and management in a low level,ignoring the analysis,association and mining cases.Deep learning is a burgeoning part of machine learning,which can make the calculation model with multiple processing levels to learn the representation of abstract data with multi levels.It can imitate human brain to understand the video data,which makes computer process massive traffic video data intelligently possible.In this background,a general method for real-time vehicle type matching based on deep learning is proposed in this paper.Based on the deep learning method,the traffic surveillance video is analyzed intelligently and the matching results are obtained.Firstly,the deep learning method is applied to the feature extraction,and the SSD network is improved for detecting vehicle and extracting all the features in the image at the same time.In addition,a new convolutional neural network named VTM-CNN(Vehicle Type Matching Convolutional Neural Network)is designed to match the vehicle type which improved the recognition accuracy of vehicle type matching method.However,the speed of deep learning is slow because of the huge consumption of GPU resources.The growth of cloud computing and big data technology can parallelize the deep learning method which make it more possible for computer processing big video data intelligently.In this paper,the Storm real-time processing platform is established.This platform realized the parallelism of vehicle type matching method based on deep learning,and the real-time requirement of vehicle type matching is achieved.Based on the characteristics of deep learning consuming a large amount of GPU resources,the scheduling algorithm in Storm is improved to make the resource allocation more reasonable and the system running more efficiently.At the same time,the real-time vehicle type matching method based on deep learning is transplanted into the embedded cluster.The evaluation shows that the constructed Storm real-time processing platform has a good fault tolerance and the real-time vehicle type matching method based on deep learning can effectively deal with continuous video data,which has a good performance and portability.