面向部件遮挡补偿的车辆检测模型

Part-oriented occlusion compensate model for vehicle detection

目的复杂场景中多目标间的遮挡,会造成车辆视觉信息损失,致使车辆检测出现漏检问题。方法为解决遮挡车辆漏检问题,提出一种遮挡补偿模型,分析车辆部件的单视点/多视点可见概率,弥补已有基于部件的车辆检测模型对遮挡区域信息描述的不足。首先,通过外观模型估计车辆候选区域,确定车辆各部件的位置和相似程度,判定车辆部件的遮挡情况,并获得外观项和结构项;其次,计算车辆区域的单视点可见概率和多视点可见概率,并获取被遮挡的部件中心点对应的单视点/多视点可见概率,作为车辆检测的补偿项,调整遮挡部分的检测得分;最后,将车辆检测的外观项、结构项和补偿项,统一到遮挡补偿模型中,实现对候选区域的车辆判断。结果实验结果表明,对比于现有的车辆检测模型,本文算法在PASCAL、MSRC以及真实场景中车辆检测结果对应的P-R曲线性能更佳。结论该遮挡补偿模型在保持虚警率的同时,能够有效提升遮挡车辆的检测准确性。

Objective The imaging conditions in real scenes are complex. Vehicle detection therefore involves many challen- ges, among which the occlusion problem is one of the most significant ones. In the object detection literature, a deformable part model applicable to rigid object detection is one of the most practical part-based models. However, this model is limit- ed by multi-object occlusion. This phenomenon results in many false negatives with a low detection score in vehicle detec- tion because of the loss of visual information under real clutter. Method To address this problem, an occlusion compensa- tion model is proposed in this study. This model analyzes the visible probability of parts according to a single viewpoint or to multiple viewpoints to compensate for the insufficiency of part-based models and thus avoid undetected errors. First, the position and similarity of each part are estimated with an appearance model in candidate regions to determine which part is under occlusion and to obtain the appearance and structure score of the object, which may be lower than the normal ones for multi-object occlusion. Second, the visible probability of a single viewpoint only considers occlusion conditions. By con- trast, the visible probability of multiple viewpoints presents occlusion states from other components that are calculated to ob- tain the compensation score for occlusion, which refines the detection score of the occluded regions. Finally, we composite the appearance, structure, and our compensation score into an integrated model to reduce false negatives. Two parameters are important in this phase, namely, part detection threshold and occlusion compensation weight. The former is applied to determine part occlusion, whereas the latter is aimed toward a refinement with the detection score of an occluded object. A high part detection threshold produces a high occlusion compensation score that in turn leads to a high false alarm rate. The condition in occlusion compensation weight is similar. They are therefore carefully selected to control the false alarm rate and decrease undetected errors. Result Visible probability modeling based on a single viewpoint is suitable for simple ca- ses, in which the visible probabilities at the same level of height are identical. On the contrary, the visible probability based on multiple viewpoints is true for complex cases, in which the visible probability near the visible part is high. The validation, which is qualitatively and quantitatively evaluated with precision-recall curves, is efficient in three data sets. The two data sets are highlighted in popular PASCAL and MSRC, and the remaining data set comes from a real scene.. Conclusion Experiment results show that our model could preserve the false alarm rate and improve the accuracy of vehicle detection under occlusion compared with the state-of-the-art model.