Abstract
We propose a new video manifold learning method for event recognition and anomaly detection in crowd scenes. A novel feature descriptor is proposed to encode regional optical flow features of video frames, where quantization and binarization of the feature code are employed to improve the differentiation of crowd motion patterns. Based on the new feature code, we introduce a new linear dimensionality reduction algorithm called “Spatial-Temporal Locality Preserving Projections” (STLPP). The generated low-dimensional video manifolds preserve both intrinsic spatial and temporal properties. Extensive experiments have been carried out on two benchmark datasets and our results compare favourably with the state of the art.