Recognition Of The Vehicle And Rear Light Videos Based On Deep Learning

With the development of China's economy,the number of urban vehicles has increased year by year,resulting in increasingly severe urban traffic pressure,which has raised more needs for the vehicle's Advanced Driver Assistance System(ADAS).Most of the communications between vehicles in traffic are fulfilled based on the signals of vehicle lights,especially the taillights.Therefore,real-time detection and accurate classification of the taillights can help to reduce the traffic accidents,which might be caused by the driver's negligence of the taillights.ADAS can capture the dynamic environment information around the vehicle through the on-board camera,and achieve the detection and recognition of the objects in video.Moreover,it can analyze and judge the road condition from the detection results,notify the driver of the road condition information in time,and assist the driver to drive safely,which has important research value and a wide range of application scenarios.However,in the real-world traffic environments,that using video image information to detect real vehicle and taillight areas and recognition the type of taillights remains a challenging problem.The previous research on the taillight recognition algorithm is mainly based on the characteristics of the color and shape of the taillights and the statistical learning method to realize the classification of the taillights.The shortcoming of these algorithm is that under natural conditions,such as light intensity and weather changes and shooting angles,the feature extraction will be greatly affected,so the detection and recognition accuracy is relatively low.With the advent of deep learning methods,the automatic extraction and representation of advanced image features by convolutional neural networks greatly improves the accuracy of target detection,and has outperformed humans in some specific task areas.However,most of the current deep convolutional network structures are complex,unable to balance speed and accuracy,and result in poor effects on unbalanced and small object detection.And it is difficult for them to analyze and identify the vehicle and the taillight in real time.Through analysis existing algorithms of vehicle and taillight detection,the major contributions of this thesis are summarized as follows: collect and produce real road traffic dataset;use YOLOv3-tiny network as the basic framework for detecting vehicle and taillight recognition.;calculate the size and the number of anchor boxes that match the best dataset;increase the number of prediction layers to better meet the detection of large,medium and small objects;increase the SPP pyramid pooling network structure so that it can freely change the size of the training set and reduce overfitting;inspired by the focal loss function,classification loss function in YOLOv3 algorithm is modified to further improve the classification accuracy for unbalanced classes.Finally,the experimental results of the proposed method in this thesis show that,the average precision of the brake light detection reaches 89.18%,the left turn signal is 88.14%,and the right turn signal is 89.69%.Through experimental comparison and analysis,it is proved that the proposed algorithm can meet the requirement of stability,which is crucial to real-time and continuous detection of vehicle lights in videos.