Real-time Traffic-light Detection Based On Depth Model

For automatic driving,the perception of the surrounding environment of the car is one of the most basic prerequisites.The current scheme mainly relies on three perception modules:lidar,millimeter wave radar and camera.The perception of road environment based on vision is a very important and irreplaceable part.Traffic signal lights are very important information in real-time video captured by vehicle cameras,which can quickly and accurately detect traffic signal lights travel safety.As the pixels of traffic lights are relatively small in the shooting pictures of vehicle-mounted cameras,the detection effect of general purpose target detection detectors is not ideal.Based on Faster R-CNN and from the point of view of feature fusion,this paper adjusts and improves the network structure by using feature fusion of the proposed area after extracting ROI_Align,which makes the detector have better detection effect on small targets such as traffic lights.A number of comparative experiments were conducted on the TT100K data set and the experimental results were analyzed scientifically and fully,which proved that our new model could effectively improve the detection accuracy of traffic lights.At the same time,the model is tested on LISA dataset,which proves the generalization performance and robustness to different scenarios.In addition to the research of region of interest feature fusion,this paper also optimizes the detection process of traffic lights in video,proposes that the detection method of subsequent frames can be determined according to the current detection results,and formulates the conditions and methods of its strategy selection,which greatly improves the detection time.In order to support and implement such a detection method,this paper also designs and implements a dual task model of target detection and location for regions of interest and a prediction algorithm for regions of interest in the next frame based on motion estimation.Finally,experiments are carried out on the video to verify the feasibility of the algorithm.