基于遗传优化与深度学习的交通信号灯检测

Traffic Light Detection Based on Genetic Optimization and Deep Learning

交通信号灯检测是先进驾驶辅助系统的关键技术之一,也是无人驾驶车辆车载环境感知的重要研究方向。本文中针对通用物体检测算法不适合信号灯这类小物体的检测和缺乏实时滑动窗口检测算法的问题,提出了一种交通信号灯检测方法,包括基于遗传优化的交通信号灯候选区域生成方法,和基于深度神经网络的信号灯定位与分类方法。其中,作为本文中研究重点的候选区域生成方法又分3部分:信号灯共用特征区域提取、基于重要性采样的信号灯候选区域参数采样和基于遗传算法的信号灯候选区域参数优化。与现有的信号灯检测方法相比,本文中所提出的方法可对横竖排的红色、绿色和圆形、箭头形信号灯进行有效检测和分类。在公开的交通信号灯数据库的对比实验表明,该方法对交通信号灯的召回率高,且能有效区分不同类别的信号灯。

Traffic light detection is one of the key techniques of advanced driver assistance system and an important research direction for the on-board environment perception in autonomous vehicles.In view of the problem that general object detection algorithms are not suitable for small objects like traffic lights and the lack of real-time sliding window detection algorithms,a novel traffic light detection method is proposed in this paper,covering the generation of the candidate regions of traffic lights based on genetic optimization and the locating and classification of traffic lights based on deep neural network,in which the former,as the focus of the study,includes three parts:the common feature region extraction of traffic lights,the parameter sampling of the candidate region of traffic lights based on importance sampling and the parameter optimization of the candidate region of traffic lights based on genetic algorithm.Compared with existing traffic light detection methods,the method proposed can effectively detect and classify the horizontal rows and vertical columns of traffic lights with red and green colors,and round and arrow shapes.The comparative experiments on a public traffic light dataset show that the proposed method has a high recall rate for traffic lights,and can effectively distinguish different types of traffic lights.