中文摘要：

针对单目视觉下复杂室外场景及人体非刚性变化多样时行人检测效率及准确率较低的问题，提出了一种基于双目立体视觉的行人检测方法．首先采用非局部视差聚合匹配方获取稠密视差图；然后以粗细两阶段分割策略获取感兴趣区域，使用最小面积阈值获取潜在行人最小尺度；结合行人最小尺度自适应地减少检测所需数据量，最后基于聚合积分通道模型完成数据建模与行人检测．实验结果表明，该方法能有效利用双目视觉信息解决单目视觉下虚假干扰问题，在实现较高检测精度的同时有效提高检测速度．

英文摘要：

To solve the problems of low detection rate and accuracy under complex outdoor scene and diverse humannon-rigid changes in monocular vision,a pedestrian detection algorithm based on binocular stereo vision was proposed.First,the dense disparity map was implemented using non-local cost aggregation method.Then the regions ofinterest was acquired by two-stage segmentation method,together with a minimal potential pedestrians scale basedon minimum area threshold.The minimum scale was used to reduce the amount of data required for the pedestriandetection,and finally,the aggregated channel features model was developed for pedestrian training and detection.Experimental results show that the proposed method can improve detection accuracy by efficiently using binocularvision information to avoid false alarms in monocular vision,and accelerate detection rate simultaneously.