| Vehicles on the road rely on the vehicle taillight language to communicate their driving status,but drivers often ignore the taillight language and cause rear-end accidents.Therefore,this paper proposes a robust method for detecting vehicles ahead and recognizing their taillight language,which is accomplished in the following five steps:First,data set acquisition,pre-processing and labeling.Video data acquisition is performed using a monocular camera under different urban road conditions,and the original image data is obtained by segmenting the video data at a frequency of 1 frame per second.The image data are de-weighted and sharpened using perceptual hashing algorithm and geometric mean filtering algorithm respectively;the vehicle head and tail area information in the image are manually labeled;the labeled tail area information is extracted and the left and right taillight area information in the vehicle tail area is manually labeled;at the same time,the left and right taillight area information in the vehicle tail area is labeled whether the vehicle tail area is lit or not;thus we get vehicle detection,vehicle vehicle detection,vehicle taillight region detection,and whether the vehicle taillight region is lit or not.Second,based on Yolov3 target detection network optimization.Whether the vehicle head tail area detection,vehicle taillight area detection and whether the vehicle taillight area is lit recognition all belong to the target detection category,this paper optimizes the Yolov3 target detection recognition algorithm,mainly for the original network in the backbone feature extraction network,loss function,default box,optimizer four parts to optimize.Thirdly,the research on the recognition of the taillight illumination of the vehicle ahead.By analyzing the characteristics of vehicle taillight language,the color space method is used to identify whether the front left and right taillight regions are lit or not;at the same time,the optimized Yolov3 target detection network is used to train the vehicle taillight region whether the data set is lit or not.Fourth,the research on the recognition of the taillight language of the vehicle ahead.Using the recognition result of whether the taillight area of the vehicle ahead is lit or not,combined with the different light language flashing frequency of taillight,five different light languages of the vehicle ahead left turn,right turn,brake,double flashing and normal driving are studied.Fifth,experimental analysis.Through six experiments to complete the thesis involved in the selection of hyperparameters and model robustness verification;in turn to complete the front vehicle detection,front vehicle taillight area detection,front vehicle taillight illumination recognition,front vehicle taillight semantic recognition.Under urban road conditions,the AP of front vehicle head area detection is: 85.87%,and the AP of front vehicle tail area detection is:90.97%.The recognition accuracy of left turn of front vehicle: 92.2%,the recognition accuracy of right turn of front vehicle: 91.2%,the recognition accuracy of brake of front vehicle: 89.6%,the recognition accuracy of double flash of front vehicle: 90.67%,the recognition accuracy of normal driving of front vehicle: 93.1%.The experimental results show that the forward vehicle detection and its taillight light language recognition system proposed in this paper have excellent robustness under urban road conditions. |
