Abstract: Multi-object detection under the smart traffic surveillance system is an important and challenging task. This paper presents a multi-object detection algorithm based on multi-orientation saliency subregion voting by confidence. Four-orientation saliency feature pixels are firstly extracted by Gabor filters. The valid subregions and scales of each object are further confirmed in terms of the multi-threshold criteria. And then valid intersection subregions and intersection states are confirmed. The duty ratios of intersection area and the relative duty ratios of valid intersection subregions are weighted and integrated to acquire the confidence vote coefficient for every subregion. The optimal detection is finally derived according to the maximum confidence vote. Experiments conducted on real traffic image sequences demonstrate that the proposed method is robust and accurate in multi-object detection under mutual disturbance and occlusion among adjacent objects, object shadow, non-uniform illumination and similar color disturbance in the background etc.

Key words: multi-object detection, multi-orientation saliency subregion, intersection state judging, duty ratios of intersection area, voting by confidence

摘要： 多目标检测是智能交通监控系统的一项重要和挑战性的工作。提出一种基于多方向显著子区域置信表决的多目标检测方法。利用Gabor滤波提取目标四方向显著特征像素；根据多阈值条件确定每个目标的有效子区域及尺度；判断有效交集子区域和交集状态；加权融合交集面积占空比和有效交集子区域相对面积占空比，获得各个有效交集子区域置信表决系数；根据最大置信表决系数获得目标最优检测。采用实际交通场景采集的视频图像序列进行实验，结果表明该方法在邻近目标交互干扰和遮挡、目标阴影、非均匀光照及背景中相似色彩干扰等情况下，均具有鲁棒、准确的检测性能。

关键词: 多目标检测, 多方向显著子区域, 交集状态判断, 交集面积占空比, 置信表决