Based On The Volume Characteristics Of Human Motion Detection

This paper introduces a novel volumetric features framework for analyzing video. Applying this framework on the video's optical flow, we learn action detectors that can detect actions in videos in real time.Motivated by the success of Harr features in static images, we extend the notion of 2D features to 3D spatial-temporal volumetric features. To maintain computational efficiency and build a real-time detector, we generalize the notion of integral video similar to integral image.In order to perform action detection on video sequences in real time, we use AdaBoost to train a classifier. Instead of building models of human body, AdaBoost uses a large count of samples to train a cascade classifier, which can perform detecting in real time. This action detector can also deal with the problems caused by traditional action detection method based on interest points.It proves that the action detectors we trained could efficiently scan video sequences in space and time, and it can perform event detection on video sequences in real time with a higher detection rate, especially for the video with simple backgrounds. It is also robust to changes in viewpoints, scales and action speed.The detectors perform well on the standard human action database-KTH.