Traffic Object Detection And Recognition Based On Deep Learning

Recently,with the increasing amounts of highways and growing complexity of traffic networks,intelligent transportation systems(ITS)have received more and more attention from both the research the society.The traffic obj ects detection and identification is one of the most important technologies and foundation for ITS.The main obj ects involved in this area includes vehicles,pedestrians,traffic signs,traffic lights and so on.This thesis mainly studies four aspects of ITS:traffic sign detection,license plate recognition,traffic anomaly detection in surveillance video and traffic density estimation.In this paper,we use the CNN-based method to detect traffic signs.Considering the difficulties in traffic signs detection,we utilize the residual connection within network and on-line hard negative example mining(OHEM)to improve the performance of Faster R-CNN.Our model finally reached 95%on the traffic sign dataset of our laboratory.For license plate recognition,semantic segmentation is implemented in our system to locate plate license,which help us to locate plate license more accurately and simplify the following jobs.We creatively implemented the segmentation network to detect license plate.Excitingly,the model designed for pixel-wise classification accurately identified the pure license plate area,which could be a rectangle or a irregular quadrilateral.Then,we transformed the quadrilaterals into rectangle license plate images using the convex hull and perspective transformation.On top of that,followed by a CNN model,all license plate characters were predicted end-to-end regardless of whatever the shape of the original license plates images.This innovative pipeline significantly improved the recognition accuracy to 98%,especially in hard cases,and reduced the time cost from 8fps to 50fps at the same time.For traffic anomaly detection in surveillance video,we first analyze the nature behind all abnormal vehicles,all broken or stalled vehicles will stay in the video background for a long time.Therefore,we use the background modeling algorithm to extract the video background frame,and use the Convolutional Neural Network(CNN)-based vehicle detection algorithm to detect potential abnormal vehicles in the background image.Finally,the decision module is designed to characterize the vehicle anomaly attribute.In the NVIDIA AI CITY Challenge competition,our proposed system achieved the 2nd with a score of 0.8645.The traffic density estimation method is based on the Convolutional Neural Network(CNN)density map estimation method.The network is used to predict the density map and the total number of vehicles at the same time,because multi-task learning could drive the model to learn feature from both macro and micro perspective at the same time.We also propose a method to generate density map according to the annotation data for vehicle detection.Compared with the conventional detection-based vehicle counting method,the density map based method achieves the better accurate rate of 95.2%and reduces the running speed to 17fps.