Enhancing Latent Features for Unsupervised Video Anomaly Detection

Recently, memory-augmented autoencoder has played a vital role in unsupervised video anomaly detection. The memory module is used for recording the prototypes of normal data and suppressing the feature representation capacity for anomalies. However, the feature representation capacity for normal data may also be suppressed due to the limited memory ability of the memory module. This may bring a serious problem to normal data being judged as anomaly. To this end, we propose a simple but effective method called Enhancing Latent Features, i.e. ELF, which integrates two modules Single Grid Feature Enhancement (SGFE) and Local Context Feature Enhancement (LCFE). They are performed sequentially to enhance the latent feature of normal instance based on single grid information and local context information, respectively. To capture the meaningful local context information, a learnable Local Sampling Location Prediction Network (LSLPN) is embedded into LCFE to predict the valuable sampling locations of local context information. Experimental results on standard benchmarks demonstrate the effectiveness of our method.