An improved two-stream CNN method for abnormal behavior detection

In recent years, abnormal behavior detection has become an active research field in computer vision and image processing. Several methods based on traditional two-stream CNN or 3D-CNN have been proposed and successfully applied in abnormal behavior detection. However, the abnormal behavior data is rarely less in the real scenario, in other words, the data is imbalanced, which may lead to overfitting problem and affect the final results. In this paper, an improved abnormal behavior detection method based on a two-stream CNN method was proposed to address the problem mentioned above. In the proposed method, DenseNet is adopted to extract both spatial and temporal features, and focal loss is employed to alleviate the influence of imbalanced data. The experiment results show that the proposed method provides a good performance in the real-world scenario.