Automatic Recognition And Semantic Analysis Of Traffic Lights Based On Vehicle Vision

Traffic lights recognition is an important supporting technology for smart cars to perceive the traffic environment,which is conducive to improving traffic safety at intersections.It has been a research hotspot that has attracted much attention in recent years.With the continuous advancement of research,the existing algorithms can better solve the problem of traffic lights recognition in the traffic environment of simple intersections.However,when the traffic environment at the intersection is more complicated or the recognition accuracy is higher,how to improve the robustness of the algorithm is a common challenge faced by current research.In view of this,based on vehicle vision technology,this paper first optimizes the threshold selection method of traffic lights self-features(color and geometric)and the recognition method of separated arrow lights,and then discusses the complex situation of strong similar interference and traffic lights being slightly blocked in detail,so as to improve the accuracy and antiinterference ability of traffic lights recognition algorithm.The main work of the full text is as follows:(1)Carry out real-vehicle driving experiments and use the driving recorder to obtain videos and images,build traffic lights videos and image materials that meet the research needs of the subject,and lay the foundation for image analysis and algorithm design.(2)Calculate the HSV color distribution law of traffic lights images,and get the color segmentation threshold under the 3σ principle(99.75%)distribution probability;derive the deformation mode of traffic lights images under real driving conditions,and clarify the geometric shape filtering thresholds under different visual field conditions;discuss the geometric constraint characteristics of the separated arrow lights,and propose a specific matching combination algorithm to reduce the miss-detection rate of the arrow lights.On this basis,a candidate region extraction algorithm based on color feature segmentation and geometric shape feature filtering is proposed.Experimental results show that the proposed algorithm has a low miss-detection rate and a better proportional control of residual interferences.(3)A static traffic lights recognition algorithm based on two-level SVM is proposed for simple situations.The algorithm firstly distinguishes the candidate regions based on the Ⅰ-level SVM and quickly filters out the interference,and then accurately verifies the preliminary judgment results based on the Ⅱ-level SVM.Experimental results show that the overall accuracy of the proposed algorithm is 99.05%,the miss-detection rate is 1.04%,and the timeconsuming is 38.1ms,which has high accuracy and real-time performance.(4)A dynamic traffic lights recognition algorithm based on spatio-temporal union is proposed for complex situations.The algorithm is based on the lighting law of the lamp body image in the time sequence to match the lighting sequence and spatial distribution,and distinguish the traffic light from strong interference based on this.Experimental results show that the overall accuracy of the proposed algorithm is 93.73%,and the miss-detection rate is0.89%.It has a certain degree of reliability and conforms to the design principle of the algorithm based on safety.(5)On the basis of traffic lights recognition,all traffic lights are grouped according to spatial location distribution and area constraints,and then all grouping results are dynamically semantically parsed to better realize early warning.