Semi-supervised anomaly detection in video surveillance by inpainting

Video anomaly detection (VAD) task can be addressed as a semi-supervised learning problem as datasets are highly biased towards normal samples. The popular reconstruction methods train the network based on only normal images. These methods detect anomaly events by comparing the input with the reconstructed image, assuming that the network can not accurately reconstruct anomalous regions. However, these methods suffer from the serious flaw that the anomaly regions are generalized sufficiently well. This problem reduces the ability of anomaly detection by narrowing the gap between reconstructed and anomaly input images. By converting the reconstruction process into an inpainting process, we introduce Inpainting-GAN, a distinctive anomaly detection model that utilizes generative adversarial networks (GANs) as the backbone network. Each input image is partially occluded in random way by a series of disjoint masks before being fed into our model, any abnormal areas in the frame will definitely be covered. Our network is trained to restore those masked regions to normal parts, which overcomes the drawback of reconstruction models. To address the temporal continuity in videos, we have incorporated optical flow as a motion constraint during the training of generator. Both qualitative and quantitative experimental results show that our proposed method shows competitive performance in VAD compared to some reconstruction methods. The experimental results on the UCSD Ped1, Ped2 and CUHK Avenue datasets have obtained AUC values of 78.6%, 94.7% and 81.4%.