Research On Deep Learning Algorithm For Traffic Image Scene Understanding

With the rapid development of economy and technology,the number of cars in China is increasing largely.Vehicle transport brings great convenience to daily life along with great stress on traffic management.Owing to the artificial intelligence,the concept of intelligent transportation system and automatic drive has quickly entered the social horizon,which needs to understand the actual traffic scenes.Related algorithm requires efficient implementation that suits the limited computing resources.In view of above issues,this thesis carries out the research on deep learning algorithm for traffic image scene understanding.The main work and innovation points of this thesis are as follows:(1)Convolutional Neural Network(CNN)shows strong learning ability in the field of image scene understanding.It is very suitable for intelligent transportation systems and unmanned vehicles.This thesis analyzes the basic structure of CNN,and reviews several classical neural network models.Meanwhile,the current problems and difficulties is summarized by expounding current CNN-based object detection algorithms and semantic segmentation algorithms.It points out the difficulty to achieve high precision and real-time speed simultaneously in algorithm platform for traffic scene understanding.In addition,some potential approaches for improving the performance are suggested.(2)The principle of YOLO algorithms is systematic described.An improved detector called TF-YOLO is carried out for the insensitivity of small and medium-sized target detection in yolov3-tiny algorithm.It aims to improve the accuracy of the algorithm and the poor detection accuracy of small target.Specifically,it introduces three strategies,clustering the prior box in the data set,multi-scale feature fusion and multi-scale prediction.Clustering on data sets could achieve a good initial value in training stage,so that training process can be carried out better,thus leaving a promoting effect on the network performance.Multi-scale feature fusion refers to the idea of DenseNet network,so that features in different layers can fusing better for approving results.Based on the FPN network,this thesis adopts multi-scale prediction to detect objects in different sizes,thus improving the detection performance of the algorithm.Finally,experiments on KITTI dataset verifies the performance of TF-YOLO.(3)Algorithms for the traffic image scene understand run in the limited hardware resources.It is difficult for YOLOv3 in real-time performance on embedded devices,despite it has an high detection accuracy based on Darknet53.To solve this problem,this thesis focuses on the network model compression algorithm based on the understanding of traffic image scene,and applies the model pruning technology to compress the network model of YOLOv3 algorithm.Experimental results show that it greatly improved the running speed of the algorithm without losing too much detection accuracy.