基于车载的运动行人区域估计方法

Locating moving pedestrian from running vehicle

针对传统行人检测系统采用全局模板搜索匹配识别易造成大量的盲目空间搜索,降低了行人检测的实时性,且难于直接服务于车载环境下的无人驾驶、行人安全保障技术等应用的问题,基于运动行人与车载移动背景的运动差异性,提出了基于光流聚类的行人区域快速估计方法。该方法首先利用局部与全局光流相结合的计算方法,获取图像光流场;根据移动背景与前景的运动差异,建立了光流聚类算法和背景估计方法,获取剔除背景的可能前景光流图。然后利用图分割算法,实现光流矢量图分割,获取运动物体候选区域。最后根据人体形态特性,判别运动物体候选区域是否为行人有效区域,从而实现行人区域估计。基于JLU-PDS和Daimler国际行人共享测试库进行试验,结果表明本文算法具有较好的检测效果,能够极大地降低行人识别的空间搜索范围,为车载主动行人保障技术、无人驾驶、智能车辆等研究和应用提供了技术支持。

Traditional pedestrian detection system uses global template for search and recognition,resulting in large useless search and reducing the realtime pedestrian detection performance.In addition,it is difficult to apply to unmanned vehicle and pedestrian safety technology in the traffic environment.Based on the difference of motion between pedestrian and vehicle moving background,a fat pedestrian area estimation method is proposed using optical flow clustering.This method uses the combination of local and global optical flow to obtain the optical flow field.According to the difference between background and foreground,the optical flow clustering algorithm and background estimation method are established.Then,the image segmentation algorithm is used to realize the segmentation ofoptical flow vector.Finally,according to the characteristics of the human body,we can judge whether the candidate region of the moving object is an effective area for pedestrians and realize the pedestrian area estimation.Experiments are carried out based on JLU-PDS and Daimler international pedestrian sharing test library,and experimental environments include the crowed,shelter,streets,fields and other auxiliary working conditions.Results show that the algorithm has good detection effect and greatly reduces the sear space of pedestrian recognition.In addition,it provides technical support and basis for research and application of vehicle active pedestrian protection technology and unmanned intelligent vehicles.