Research On Structured Object Detection With Occlusion Reasoning

Object detection is the basic task in the video surveillance system, and provides effective decision information for high-level tasks. However, occlusion frequently occurs among multiple objects in complex traffic environment, which leads to the appearance loss for partial object region and causes increment of undetected error. The main issue for occluded generic object detection is the analysis of occlusion relationship, which also becomes the wide consensus to improve undetected problem.In the beginning, the overview of handling occlusion object is discussed to illustrate the problem of existing methods. Then, occlusion compensate model is proposed by combining part-based model and occlusion relationship analysis methods. Original (occluded) object is parsing into multi-parts with structural representation. Basis appearance and structure information is provided from part-based model. In order to recover the loss of visual information, part location and object size is used to generate object visibility based on single clue or multiple clues, which is also considered as occlusion compensation information to complete occluded object detection. This thesis includes these following contents:(1) There are three problems in research status about occluded object detection, such as occlusion probability estimation, occlusion type estimation and occluded part analysis. And new basic hypothesis of occluded object detection model is proposed that parts visibility could be used to estimate object occlusion relationship based on part-based model.(2) Several key issues are discussed in this thesis, including the motivation and definition of object structural representation, the learning and matching procedure of object templates. Then the limitations of object structural representation is pointed out that undetected problem still remains due to failure estimation about object appearance under object partial occlusion.(3) To solve undetected problem, the determination of part occlusion label is introduced to part-based model, then part-oriented occlusion compensate model for object detection is constructed to realize occluded object detection. Part visibility estimation is calculated as part visual probability based on single clue or multiple clues, which is further used to generate the score of occlusion compensation to improve the performance of object detection under occlusion.(4) To reduce the influence of false alarms caused by occlusion compensation information, the compensation reliability is analyzed with different parts, and the compensation term is modified by weighting strategy. Including qualitatively and quantitatively with precision-recall curves, the validation of weighting model is shown on PASCAL datasets. Compared to the state of the art, the experiment results show that our model could preserve false alarm rate and improve the accuracy of vehicle detection under occlusion.