Anomaly Detection via Midlevel Visual Attributes

Figure 1

Overview of our approach. This is an example of two-level spatiotemporal pyramid. The input is a video stream. Then a 3D volume around a pixel is constructed represented by the outer red cube. Then it is segmented into 8 () smaller cubes denoted by different numbers in this figure. The smaller cubes form the lower but finer level of the pyramid. HOG features are extracted for each smaller cube. And the HOG features of upper level cube can be constructed efficiently from lower level cubes. We use visual attribute representation to bridge the semantic gap between low-level feature and high-level event. The three-level (feature-attribute-event) framework can be modeled by extreme learning machine. Finally the anomaly detection is completed by combining the outputs of the machine.