Traffic Light Detection And Recognition Technology Research Based On The Image

In the driverless era,unmanned vehicles use various types of sensors to enable environmental perception and integrate the perceived information to achieve decision making and planning.With the development of driverless technology,traffic light detection and recognition technology is one of the core technologies to implement driverless.Timely and accurate recognition of traffic lights will effectively improve the safety of vehicles,ensure traffic safety and implement vehicle-road cooperation.Based on the traditional image processing algorithm to complete the traffic light detection and recognition.The algorithm process is to determine the area to be detected by using the light color features,geometry,and other characteristics of the traffic light,and then extract the features to complete the recognition of color.This algorithm is easily affected by the surrounding environment.In this thesis,we propose deep learning networks based on detection and recognition double steps to solve the problem of low recognition rate of single detection network.Using tracking,the accuracy of detection and recognition is further improved to obtain the forward traffic light recognition within0-100 m.The main contributions of this thesis are as follows.1.Confirm the networks for solving the two tasks of detection and recognition,and train the networks that are more fit for domestic scenarios.The BDD100 k dataset was used to confirm the better model for small target detection.The mean average precision(m AP)of YOLOv5 model is highest up to 55.4% for comparison.In the first step,after pretraining the YOLOv5 model,the self-researched dataset with different environmental conditions was used several times to realize finetune,and the m AP reached 64.6%.In the second step,the traffic light shape is segmented,and the Resnet18 network is trained on the DTLD dataset to achieve recognition with m AP reaching 95%.2.Combine two network models for detection and recognition and test them using different datasets to confirm the combined model with better generalization capability.The detection candidate boxes are reduced to determine the unique match of the projection boxes based on the distance and confidence scores.The detection box is obtained and fed into the recognition model to complete the recognition.The result measures 99.2% of m AP for 0-70 m and 83% of m AP for 70-100 m.3.Design the tracking module to enhance the stability and improve the effectiveness of the model.Problems such as false detection of detection box,false detection of color state,and missed detection in single frame visualization.Problems such as unstable detection and recognition results in continuous frames visualization.In the tracking module,the corresponding algorithms are designed to solve the specific problems.Based on the historical detection boxes,the Kalman filter is used to complete the tracking correction.Based on the historical color state,the light uses color confirmation algorithm to complete tracking color fix,and the black light adds circle detection to complete tracking color fix.The color states of the traffic lights in the same frame are voted to ensure stable output.