基于数据流和精确定位的多线程行人探测系统

Precise positioning multithreading predestrian detection system based on data stream

行人防撞警告系统(Pedestrian Collision Warning System,PCWS)是车载主动安全系统的一项主要功能。目前通常的行人检测算法在商用化过程中的主要制约是庞大的计算量导致的低处理帧率。研究了PCWS中的关键技术,综合运用了基于VP评估的空间限制滑动窗口分布、基于数据流的多线程处理流程、基于CENTRIST特征的快速线性SVM分类器、基于直方图交叉核SVM(HIKSVM)的非线性分类器等多种加速技术,达到了实时处理的要求。对于CENTRIST特征不能精确描述对象轮廓所造成的定位不准问题,通过基于高斯权重分布的极大化抑制以及基于外轮廓边缘差异微调包围框尺寸来精确重定位探测框,以满足商用化对测距准确性的要求。

Pedestrian collision warning system （PCWS） is one of important functions of on-board active safety system. The main restriction in the process of commercial applications is large amount of calculation. This reduces the processing frame rate. This paper research the key algorithms of the PCWS, such as space limit distribution of sliding windows based on estimation of vanish point, muhithreaded processing based on data stream, rapid linear SVM classifier based on CENTRIST feature, histogram intersection kernel SVM （HIKSVM）. Comprehensive the above techniques, has reached the requirement of real-time processing. CENTRIST feature cannot accurate descript the object contour, this lead to the detected positions are not accurate. This problem is solved by applying maximum suppression based on Gaussian weight distribution and fine-tuning bounding box based on differences between inside and outside of outer contour. Through the above methods, meet the requirements of lead to ranging accuracy.