A new method of abnormal event detection based on sparse reconstruction

We proposed a framework of abnormal event detection via sparse reconstruction cost over the normal basis. Given a collection of normal training samples, we extracted the Multi-scale Histogram of Optical Flow feature and generate a overcomplete dictionary. Then a novel online dictionary selection method with group sparsity constraint is designed, which solve the dictionary learning problem in less memory and time consume. To detect whether the frame is abnormal or not, we use sparse reconstruction cost (SRC) over the testing sample to indicate the probability to be abnormal. By considering the basis in the dictionary appears frequently in the training dataset and the cost to use it in the reconstruction should be lower. The proposed weighted SRC is more robust compared to other outlier detection criteria. Experiments on UMN dataset and comparison to the state-of-the-art methods show that our framework outperform the others.