Intensive Pedestrian Detection Algorithm Based on Key Points
Hongquan Qu, Gaoyi Guo, Yue Han · 2020
There are two problems in the intensive pedestrian detection task: a large span of pedestrian scale and serious crowded. This paper adopts the object detection of the new trend in 2019. The algorithm of anchor-free mechanism to solve the first problem, and the soft-NMS mechanism to solve the second problem. We used FoveaBox based on soft-NMS algorithm improvement. First of all, compared with the traditional anchor-based detector, anchor free detector has a stronger generalization ability and also performs better in dealing with the size changes of different aspect ratios, anchor covers a limited scale range and shows some struggling in the intensive pedestrian detection task. Secondly, we improved the general NMS algorithm by adopting a soft-NMS algorithm. Soft-NMS algorithm can effectively improve the recall rate and average accuracy in the face of crowded problems. We used the subway passenger flow in peak period as the data, and compared the FoveaBox based on soft-NMS with the generic Faster R-CNN and YOLOv3, which proved that the FoveaBox improved by soft-NMS was an effective algorithm to deal with intensive pedestrian detection tasks.