Find in Library

Anomaly detection in traffic surveillance videos is a challenging task due to the ambiguity of anomaly definition and the complexity of scenes. In this paper, we propose to detect anomalous trajectories for vehicle behavior analysis via learning regularities in data. First, we train a sequence-to-sequence model under the autoencoder architecture and propose a new reconstruction error function for model optimization and anomaly evaluation. As such, the model is forced to learn the regular trajectory patterns in an unsupervised manner. Then, at the inference stage, we use the learned model to encode the test trajectory sample into a compact representation and generate a new trajectory sequence in the learned regular pattern. An anomaly score is computed based on the deviation of the generated trajectory from the test sample. Finally, we can find out the anomalous trajectories with an adaptive threshold. We evaluate the proposed method on two real-world traffic datasets and the experiments show favorable results against state-of-the-art algorithms. This paper's research on sequence-to-sequence regularity learning can provide theoretical and practical support for pattern anomaly detection.