基于CS-SD的车载环境下实时行人检测模型

针对车辆辅助驾驶系统中行人检测的实时性问题, 提出一种基于路面边缘线标定结合显著性纹理检测(CS-SD) 的算法和定位方向梯度直方图(L-HOG) 的行人检测模型, 应用CS-SD算法替代穷尽搜索快速标定图像中的行人区域, 应用L-HOG快速提取行人特征, 并采用附加核心的支持向量机(AK-SVM) 进行高效目标分类。分析结果表明: 在个人计算机上对包含832个行人的500幅图像进行检测时, 模型正确检测720个行人, 检测率为86.5%, 误检率为4.1%, 检测时间为39ms; 在基于BF609的车载行人检测系统上对包含988个行人的48 400幅图像进行检测时, 模型正确检测861个行人, 漏检127个行人, 误检13个行人, 检测速度为20fps。可见, 提出的行人检测模型在不降低检测率的前提下, 可以达到满意的检测速度, 并且可以用于实时行人检测车载设备。

Model of real-time pedestrian detection under vehicle environment based on CS-SD

In order to solve the real-time problem in the advanced driver assistant system, a model of pedestrian detection based on the calibration of side-of-pavement line and saliency texture detection (CS-SD) and the location histogram of oriented gradient (L-HOG) was proposed.The CS-SD algorithm was used instead of exhaustive search to quickly mark pedestrian area in the image.The L-HOG was used to quickly extract pedestrian feature, and additive kernel support vector machine (AK-SVM) was used to efficiently classify detected objects.Analysis result shows that when 500 images including 832 pedestrians on personal computer are detected, the model detects 720 pedestrians correctly, the detection rate is 86.5%, the error rate is 4.1%, and the detection time is 39 ms.When 48 400 images including 988 pedestrians on vehicle pedestrian detection system based on BF609 are detected, the model detects 861 pedestrians correctly, misses 127 pedestrians and detects 13 pedestrians in error.The detection speed is 20 fps.Underthe premise of not reducing the detection rate, the proposed pedestrian detection model can reach satisfying detection speed and can be used in vehicle equipment of real-time pedestrian detection.