Recognition of Abnormal Behavior of Crowd based on Spatial Location Feature
Li Zhang, Jin Han · 2020
Recognition of anomalous behavior for crowd gathering and dispersal faces challenges such as target scale change, target short-term disappearance, and target occlusion, which can result in low recognition accuracy. We propose an algorithm based on features of spatial location relationships for the recognition of crowd behaviors to solve the problems mentioned above. We first modified the network based on YOLOV3 by removing the 32x downsampling layer and replacing it with a 4x downsampling layer to further improve the detection of small scale targets, and we re-select the anchor boxes via K-means clustering algorithm to make the network more robust in training. We then perform data association by using the Kalman filter and Hungarian algorithm for target matching between frames to continuously track pedestrians and capture the motion trajectory. The trajectories of pedestrians are transformed to calculate the features of spatial position relationships, and ultimately used as a basis for determining whether abnormal behavior has occurred in the crowd. We validate our approach on PETS2009 dataset and it achieves an 89.0% accuracy with a processing speed of 29.7FPS It is of great significance to maintain the safety of people's lives and reduce the casualties caused by emergencies in public places.