Random Forest Based Multi-View Fighting Detection with Direction Consistency Feature Extraction

Nowadays, with the increasing number of surveillance cameras, the demand for intelligent video surveillance systems is continuously growing. Detection of fight behaviors is an important and challenging research field of intelligent video surveillance systems. In real life, the camera shooting views are usually different in complex scenarios, so a multi-view approach which performs well in videos with different shooting views is critical. In order to improve the performance of existing methods in videos with different shooting views and solve the misjudgment on non-fight, such as running, talking, etc., we analyze the motion characteristics of fight behaviors and propose two features named Direction Consistency feature and Weighted Direction Consistency feature to distinguish fight and non-fight behaviors. Based on the statistics of features, we define the final feature which is fed into the Random Forest classifier. Moreover, the proposed method is evaluated with the CASIA dataset, and the results indicate that the proposed approach can improve the accuracy, missing alarm and false alarm for the detection of fight behaviors, and it is very robust against videos with different shooting views.