多车道复杂环境下前方车辆检测算法  

Front Vehicle Detection Algorithm in Multi-lane Complex Environment

作　　者：孔栋[1] 黄江亮 孙亮[1] 钟志伟 孙一帆[1] 

机构地区：[1]山东理工大学交通与车辆工程学院,山东淄博255091

出　　处：《河南科技大学学报（自然科学版）》2018年第2期25-30,35,共7页Journal of Henan University of Science And Technology:Natural Science

基　　金：国家自然科学基金项目(61074140;51508315)

摘　　要：为了解决智能车安全辅助驾驶系统中前方车辆目标的检测问题,提出了一种基于改进阴影多特征与深度网络学习的车辆检测算法。基于前方车辆与本车存在安全距离,选取道路图像底部几行作为候选道路背景并对其预处理排除干扰,通过差分得到车底阴影增强图像。利用自适应阈值法确定图像灰度分割阈值并对道路二值化图像进行形态学预处理。然后,利用最小外接矩形框选候选车辆目标,结合车底阴影几何位置特征、对称度特征进行滤波生成车辆假设。最后,基于局部二值模式纹理特征和深度学习方法验证车辆假设。实验结果表明:在复杂干扰的多车道环境中,算法可以有效地检测前方车辆目标。In order to solve the problem of front vehicle target detection in intelligent vehicle auxiliary driving system,a vehicle detection algorithm based on improved shadow feature and depth network learning was proposed.Based on the safety distance between the front vehicle and the vehicle,the bottom of the road image was selected as the candidate road background and the interference was eliminated by preprocessing.Then the bottom vehicle shadows enhancement image was obtained through the difference.The adaptive threshold method was used to determine the gray scale threshold of the image and the morphological preprocessing of the roadbinarized image was carried out.Then,the candidate vehicle target was selected by the minimum circumscribed rectangular box.The vehicle hypothesis was generated by combining the geometric characteristics of the vehicle shadow and the symmetry characteristics of the vehicle. Finally,the vehicle hypothesis was verified based on local binary pattern( LBP) texture feature and depth learning method. The experimental results show that the algorithm can effectively detect front vehicle target in the multi-lane environment with complex interference.

关 键 词：车辆检测 改进阴影多特征 局部二值模式纹理 深度网络算法 机器视觉 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]