|
To overcome the challenge of occlusion in pedestrian detection, an occlusion-aware detector (OAD) is proposed in this paper. Specifically, to deal with the dilemma problem of non-maximum suppression in solving intra-class occlusion, the crowd-counting method was designed to denote the density map of pedestrians, which can be easily applied to anchor-free models. For inter-class occlusion, we innovatively introduce contrast learning into pedestrian detection to weaken the features of occlusion and strengthen the features of visible parts. Extensive experiments were conducted on two benchmarks, including CityPersons and Caltech, and the results show that our model can achieve significant performance improvement compared with state-of-the-art approaches. |
|
Keywords:Pedestrian detection, Non-Maximum Suppression (NMS), Occlusion-aware. |
|