Robust Deep Interaction Recognition Framework with Multi-Stage Feature Analysis

Computer vision and pattern recognition have been interested in human interactions on images. It is indeed that interaction recognition is an area of interest in most research areas. If this is the case, the research proposed uses a Deep Neural Network to design an activity recognition system. An HSI color transformation is integrated at the initial stages of the system to improve the clarity of the video frames. Additionally, we apply Gaussian filters to minimize noise interference. MOT and statistical methods are used to silhouette extraction. The extraction process uses the Texton maps, FAST, and Local Intensity Order Pattern (LIOP) technique. The Gray Wolf algorithm is then used to discriminate the features, and then the most meaningful independent content that describes the data structure is identified. Finally, the ANN is fed the last feature and classified into relevant human interactions. This approach is tested with the SBU Interaction dataset and the recognition rate attained is 92% through the experimental method.