Traffic Object Detection And Recognition Based On Deep Learning

With the rapid development of artificial intelligence technology,driverless technology can be realized in reality.And the perception of the real-time situation outside the car is one of the core aspects of unmanned driving.One of the means for driverless vehicles to obtain external information is to use visual sensors to sense real-time traffic conditions,such as traffic signals,traffic signs,pedestrians and vehicles.As an important task in driverless driving,the recognition of traffic signs is extremely complex in the actual detection process.For example,the traffic signs are blocked and the line of sight is blurred in rainy days.Therefore,the recognition of traffic signs is a challenge in speed and accuracy.The recognition of traffic signs can be realized by Convolution Neural Network based on deep learning.It can detect and locate traffic signs by training and learning the characteristics of specified targets in a large number of image data.At the same time,considering the use of the model in the driverless mobile terminal,this paper selects a lightweight object detection algorithm to recognize the traffic signs,and the data set this paper used is CCTSDB.Firstly,this paper studies three classical lightweight object detection algorithms,and expounds their network structure and advantages.Then,the performance of three object detection algorithms for traffic sign recognition is verified by data set of CCTSDB and evaluation standard of m AP.The object detection algorithm of YOLOv4 Tiny is selected as the lightweight model for follow-up research.This paper proposes an improvement on the K-means method used by YOLOv4 Tiny to obtain the prediction box in detection,and points out the influence of IOU threshold between the prediction frame and the actual target on the detection results in the process of YOLOv4 Tiny model training and detection.In order to eliminate the error of Anchor mechanism in the detection process of YOLOv4 Tiny,the adjustment measures after K-means are proposed.At the same time,K-means++ is used to improve the stability of clustering.The improved Anchors improve the detection accuracy of YOLOv4 Tiny by 7.39 percentage points.Secondly,according to the characteristics of small targets in data set,this paper proposes feature fusion enhancement and adding atrous spatial pyramid pooling module to solve the problems of low resolution,fuzzy picture,less information and more noise of small targets.Using feature fusion enhancement improves the ability of the model to detect shallow information,and using atrous spatial pyramid pooling improves the perception ability of the model to different scales.The improved network structure improves the accuracy of target detection model by 1.49%.Finally,in order to solve the shortcomings of the original post-processing method of the model,this paper proposes to use soft-NMS for replacement and improvement and improves the detection accuracy by 0.58 percentage points.Based on the improved object detection model,this paper uses Py Qt5 to develop a simulation system for detecting traffic signs,and uses the simulation system to test the pictures with traffic signs randomly taken by mobile phones successfully.