Research On Traffic Light Detection And Recognition Method Based On Deep Learning

The detection and recognition of traffic lights is the key technology to realize unmanned driving and assisted driving,the accuracy,stability and timeliness of detection and recognition are directly related to the safety of intelligent driving.In complex and changeable traffic scenes,traffic lights are smaller than other objects,and appear as small scale objects with fewer pixels in the image,and are easily interfered by car tail lights,brake lights,LED billboards,and street lights,so it is difficult to detect and recognize.Traditional detection methods that first extract the ROI area based on features such as color,shape,HOG,LBP,SIFT and other features and then classify and recognize have great limitations,In recent years,deep learning has made breakthrough progress in the field of object detection,and both detection speed and recognition accuracy are far ahead of traditional algorithms.This thesis studies the detection and recognition method of traffic lights based on deep learning,and improves its deficiencies,the main work is as follows:(1)Based on the traffic lights data set Bosch Small Traffic Lights Dataset(BSTLD),the current mainstream object detection algorithms Faster R-CNN,SSD,YOLOv3 and YOLOv4 were comparative experiment,it is found that the above depth models have insufficient detection capabilities and poor robustness when detecting small scale traffic lights,and it is prone to missed detection,misdetection,and inaccurate positioning.(2)Selected the YOLOv3 algorithm with better performance in the comparison experiment as the basis and make the following improvements,to improve the model’s ability to detect small scale traffic lights:(a)Replace the feature extraction backbone network: Replaced the FC fully connected layer and Softmax classification layer of the VGG16 network with 3×3convolutional layer as the feature extraction backbone network of the YOLOv3 algorithm,to enhance the ability to describe detailed features.(b)Leapfrog feature fusion: Firstly,increase the number of DBL channels executed by the final output layer of the feature extraction backbone network from 128 to 256,then made the 13×13 feature map directly up sampled 4 times,and then perform feature fusion with the 52×52 feature map,to enhance the semantic discrimination ability of52×52 shallow feature maps for small scale objects.(c)Clustering & scaling to obtain a new prior box: Use the K-means algorithm to cluster Ground Truth in the BSTLD training set,and then use the linear scaling mechanism to discretize the clustered prior boxes to obtain new prior boxes that are more suitable for small scale traffic light sizes,and then improve the accuracy of model positioning.(d)Experiments have verified the effectiveness of the above three improved methods alone and in combination with each other in improving model detection capabilities,compared with the original YOLOv3 model,the YOLOv3＿VGG16-LFF model which replaces VGG16 as the feature extraction backbone network and uses leapfrog feature fusion performs best,and can bring about a 14% improvement in m AP.(3)Most of the current traffic lights recognition algorithms are based on motor vehicle signal lights(ie,"circular lights")and direction indicator lights(ie,"arrow lights"),and rarely involves the detection and recognition of "countdown digital lights",This thesis proposed a countdown digital light recognition method based on the combination of YOLOv3 and OCR,compared with the traditional OCR recognition method,the recognition accuracy has been greatly improved.