Motion interaction field for detection of abnormal interactions

This paper proposes a new method for the modeling of interactions among objects and for the detection of abnormal interactions in a video. To model interactions among multiple moving objects, we design a motion interaction field (MIF) that is similar to a water waveform generated by multiple objects moving on the surface of water and that describes the intensity of motion interaction in a video. Using the MIF, we establish a framework to detect abnormal interactions, which consists of rule-based decision about regions of interest and dictionary learning-based anomaly decision for these regions. The regions of interest are determined as the regions remaining after filtering out collision-free regions that are recognized clearly to be normal by a rule-based decision based on the shape of MIF. The MIF values in these regions are then used to construct spatiotemporal features for the detection of abnormal interactions by a dictionary learning algorithm with sparse representation. In the experiments, the effectiveness of the proposed method is validated through quantitative and qualitative evaluations with three datasets containing typical abnormal interactions such as car accidents, crowd riots, and uncontrolled fighting.