Occlusion handling using discriminative model of trained part templates and conditional random field

Occlusion handling is crucial for developing ADAS in urban environment. Deformable Part Models (DPM) have already demonstrated state of art results in object detection. However, they fail to handle the occlusion due to inaccurate scores for part detectors when some object parts are occluded and not visible. To handle the imperfectness of part detectors in occlusion, we propose a two layers classifier using DPM and conditional random field (CRF). We use parts contextual information and their spatial configuration from DPM to train and optimize CRF parameters. Occlusion states are defined based on visibility of parts and considered as latent variables in CRF. To learn CRF parameters, stochastic gradient descent with belief propagation is used to optimize CRF objective function for latent variables. Experimental results on recorded data in real urban environment and PASCAL VOC dataset prove the effectiveness of the proposed approach to handle difficult occlusion situations.