Increment prediction for video anomaly detection
Xinggang Guo, Bai Wang · International Symposium on Robotics, Artificial Intelligence, and Information Engineering (RAIIE 2022) · 2022
The Appearance-based and motion-based dual-stream anomaly detection algorithm uses optical flow to extract motion features in normal video, but optical flow only considers the pixel displacement relationship between two adjacent frames, while ignoring the latent semantic association of continuous changes in multi-frame images. This paper proposes an incremental prediction-based video anomaly detection algorithm (IP-VAD), which first uses an encoder to extract corresponding features of consecutive frames, and then reconstructs feature maps from appearance and motion directions, respectively. In the appearance direction, the last frame is reconstructed by the appearance decoder; in the motion direction, the feature differences of successive frames are stacked and fed into the motion decoder, which then predicts the increment of future and last frames. Finally, future frames are obtained using the reconstructed last frame and the predicted increment. During the test, the peak signal-to-noise ratio is used to measure the reconstruction quality of future frames, and the threshold is used to determine whether it is abnormal. The experimental results on the public datasets UCSD Ped2 and CUHK Avenue show that the AUC indicators of our method are improved by 1.0% and 0.2%, respectively.