A Video Abnormal Detection Framework based on Appearance-Motion Fuse Memory

Abnormal events are usually considered extraneous event deviation from standard events. Context and Movement mode play a pivotal part in determining abnormal events. Due to the forms being unbounded and ambiguous, it is impossible to straight fitting abnormal. Based on this requirement, following a recent series of methods, we used an object detector, followed by abnormal detection at the object level. This paper proposes a brand new video abnormal detection approach. By occupying Appearance-Motion Fuse Memory (AMFM), the new pattern will utilize the appearance message and motion representation. The Spatio-temporal model uses a two-stream encoder. The two streams of input refer to several preceding video frames and their corresponding optical flow data. In the course of training, the two-stream encoder takes the role of extracting the superficial and moving features by the merely normal entities. We discover that a design like this two-stream encoder has the ability to commit the characteristics of normal samples to memory. Experimental results on the commonly used ShanghaiTech dataset and UCSD PED2 dataset verified that our model achieves competitive performance. While we conduct ablation studies to prove the validity of the components constructed in our model.