基于显著性检测的实时车载行人检测方法

Real Time Vehicle Pedestrian Detection Method Based on Saliency Detection

针对车辆辅助驾驶系统中行人检测的实时性问题,提出了一种新的基于显著性检测和参数融合的LUT-HOG车载行人检测方法.在车载行人检测方法中提出基于多局部区域的空域调频对比度的显著性检测方法,快速标注图像中的"行人"区域;在"行人"区域,提出基于参数融合的LUT-HOG的快速行人特征提取算法;采用逼近线性SVM计算时间的AK-SVM进行高效分类.该算法通过使用显著性检测代替穷尽搜索办法,并用快速的行人特征提取和分类算法,能够大幅提高行人检测的速率.在Visual Studio 2012软件中使用C语言验证行人检测算法的正确性和实时性.基于INRIA数据库的测试结果表明,在不降低准确率的前提下,该行人检测系统可以达到25帧/秒(640×480)的检测速率.然后,将行人检测系统移植到BF-609的开发板上进行实时行人检测,在不使用GPU的前提下,可以达到20帧/秒的检测速度,可满足车载行人检测系统的实时性要求.

A real-time vehicle pedestrian detection method was realized by using saliency detection and parameter fusion LUT-HOG （Look Up for Table-Histogram of Orientation Gradient）. The saliency de tection based on the multi local region frequency-zone could annotate ＂pedestrian＂ zone quickly. In the ＂pedestrian＂ zone, the multi-parameter fusion L-HOG was proposed to extract the features quickly and AK-SVM （Additive Kernel Support Vector Machine） efficiently classified the objects, which has the same computational time with the linear SVM. This pedestrian detection system was evaluated with the C language in the Visual Studio 2012. The results based on the INRIA dataset show that this pedestrian detection can process 25 frames per second （640× 480） without sacrifice of accuracy. At last this system is transplanted to the BF-609 board to detect pedestrians. This vehicle system can process 20 frames per second and can be as the real time vehicle equipments without using the GPU.