Multi-view crowd congestion map generation based on ensemble learning
Kourosh Khoshelham, Yan Li, Majid Sarvi, Milad Haghani, Tian Yuan · Transportation Research Board 98th Annual MeetingTransportation Research Board · 2019
Multi-view video surveillance has been a major research area in crowd congestion management. By exploiting complementary information captured by multiple cameras, the limited views and occlusion in single views can be addressed to gain an insight into the whole monitored space. However, multi-view surveillance has been widely applied to microscopic crowds analysis, for example pedestrian detection and tracking, while macroscopic level analysis, which deals with the whole crowd, has received little attention. Level of service (LOS) is the most widely accepted standard of measuring congestion at macroscopic level and level of service maps are the most straightforward way of showing distribution and variation of congestion. The authors propose a multi- view framework for the generation of level of service maps based on an ensemble of state-of-art Convolutional Neural Networks (CNN). Several combination rules are compared and evaluated on two datasets, both in sparse and dense scenarios. The authors' results show that this fusion framework improves the accuracy of level of service map generation, from 83.2% to 89.8%, and eliminates vision of blind spots in single views. The authors' framework is implemented to a 3D geographic information system (GIS) platform, which provides a suitable interface for multi-view crowd congestion management. A loading test results show that a maximum of 48 cameras can be processed at a map refresh rate of 2 seconds.