Action Recognition In Unconstrained Videos

Action Recognition in videos is an active research field and is used widely in video surveillance, video retrieval.Still, existing systems fall short of the application’s needs in real-world scenarios. In this paper,action recognition in unconstrained videos where actors’identity and clothes, scene background and illumination, camera viewpoint and motion, and image, resolution and quality all change between the various clips of the same action.According to the characteristics of the unconstrained video, the technology of action recognition in unconstrained video is research. The main task includes background subtraction, compensate for camera movement, making the two-level based on the neighborhood of spatial and temporal features and modeling.(I):Based on Motion Interchange Patterns, we decouple image edges from motion edges using a suppression mechanism, and compensate for global camera motion by using a fitted registration scheme. We build MIP-sc feature.(II):In this paper, we build the second sub-layer descriptor is based on the neighborhood of spatial and temporal features (MIP-sc). This method compensates lost spatial and temporal information for the bag-of-words technique. We have proved it to be more effective representation of the action by experiment.(III):We have tested our descriptor on the ASLAN dataset, HDMB51 dataset, UCF50 dataset, KTH dataset. We gained good accuracy and is robust to illumination, camera viewpoint and motion.