Abstract: Abstract: The occlusion problem poses challenges to the current object detection. The presence of occlusion could destroy the overall structure of the object, which is likely to incur missing detections and false positives during the detection. Although the common methods for handling occlusion have greatly enhanced the performance of occlusion detection, there remains no specific quantitative analysis of the occlusion components and the impact of different occlusion ratios on the detection performance. In this paper, based on the data-driven method, a large number of uniform occlusion datasets were generated by simulation, named as More than Common Object Detection (MOCOD), and the detection performance under different occlusion ratios was analyzed quantitatively. On the basis of the analysis of occlusion’s influence, according to the occlusion ratios, the decay weight was introduced to select high-quality positive samples for the model training, thereby effectively improving the detection performance under occlusion conditions.

Key words: Keywords: deep convolutional neural networks, object detection, occlusion handling, occlusion datasets