Unsupervised video anomaly detection based on multi-timescale trajectory prediction

Video anomaly detection refers to the identification of events that differ from normal behavior. Most of the commonly used methods currently use the reconstruction method based on appearance features. However, appearance features are unstructured signals that are sensitive to noise, and the redundant information contained will also increase the burden of distinguishing signals from noise during training. Reconstruction methods try to minimize the reconstruction error of the training data but cannot guarantee that the reconstruction error of anomalous events is large. From this, we propose an unsupervised video anomaly detection algorithm based on multi-timescale trajectory prediction. We use the object tracking network to detect and track pedestrians in the scene and send them to the multi-timescale trajectory prediction and velocity calculation modules for training. Due to the different motion durations, we add a multi-timescale mechanism to predict pedestrian trajectories and introduce step signals in digital signal processing for trajectory subsequence segmentation. During the testing of abnormal videos, the irregular motion behavior of pedestrians cannot be predicted by the normal model and will result in higher trajectory outliers. Similarly, the velocity calculation module constrains and calculates pedestrian velocities at different camera views, and events that differ from normal velocity can be detected under the dual constraints of space and motion (time). Compared with state-of-the-art methods and other anomalous event detection methods, the proposed model has certain advantages in frame-level AUC.