Accurate Object Recognition with Assembling Appearance and Motion Information

Abstract

How to effectively detect object and accurately give out its visible parts is a major challenge for object detection. In this paper we propose an explicit occlusion model through integrating appearance and motion information. The model combines together two parts: part-level object detection with single frame and object occlusion estimation with continuous frames. It breaks through the performance bottleneck caused by lack of information and effectively improves object detection rate under severe occlusion. Through reevaluating the semantic parts, the detecting performance of partial object detectors is largely enhanced. The explicit model enables the partial detectors to have the capability of occlusion estimation. By discarding the geometric representation in rigid single-angle perspective and applying effective pattern of objective shape, our proposed approaches greatly improve the performance and robustness of similarity measurement. For validating the performance of proposed methods, we designed a comparative experiment on challenging pedestrian frame sequences database. The experimental results on challenging pedestrian frame sequence demonstrate that, compared to the traditional algorithms, the methods proposed in this paper have significantly improved the detection rate for severe occlusion. Furthermore, it also can achieve better localization of semantic parts and estimation of occluding.