Adaptive Crowd Segmentation Based on Coherent Motion Detection

In order to obtain effective crowd relationships for crowd behavior analysis, an adaptive crowd segmentation method based on coherent motion detection is proposed. This method can improve the accuracy of segmentation results and be adaptively applied to various collective scenes that have different distributions at different scales. Firstly, an orientation clustering algorithm and a spatial joint strategy are proposed to preliminarily profile all agents into several partitions with different motion orientations. Then, the Natural Nearest Neighbor algorithm is introduced to construct the adaptive crowd motion networks combining with the profiling results, which can describe the neighborhood relationships of agents with stronger coherence. Finally, the improved Coherent Neighbor Invariance optimized by fusing motion information of neighbors is proposed to segment crowds with coherent motions from the crowd motion networks. The experiment results on videos depicting real-world crowd scenes indicate that the proposed method is effective and adaptive to various scenes.