Comparison of cuboid and tracklet features for action recognition on surveillance videos

For recognition of human actions in surveillance videos, action recognition methods in literature are analyzed and coherent feature extraction methods that are promising for success in such videos are identified. Based on local methods, most popular two feature extraction methods (Dollar´s “cuboid” feature definition and Raptis and Soatto´s “tracklet” feature definition) are tested and compared. Both methods were classified by different methods in their original applications. In order to obtain a more fair comparison both methods are classified by using the same classification method. In addition, as it is more realistic for recognition of real videos, two most popular datasets KTH and Weizmann are classified by splitting method. According to the test results, convenience of tracklet features over other methods for action recognition in real surveillance videos is proven to be successful.