Research On Action Detection Technology And Implementation Of Software System

Visual based action detection is an important topic in the field of computer vision, and it has attracted much attention from researchers for its wide application prospect. This thesis focuses on action detection issues under the framework of machine learning, including the determination of the temporal and spatial position where the action occurrs, the action detection algorithm and the implementation of action detection software system. The specific research works are as follows:On the basis of summarizing and evaluation of existing moving target detection methods, an adaptive motion area estimation method which can make full use of temporal and spatial coherence of action is proposed to remove the outliers produced by inaccurate motion detection. This method is applied to improve the accuracy of motion area detection, and then the fusion feature are extracted in the motion area to ensure the accuracy of the described feature while reducing dimension of feature. Finally, SVMs are used to train detection model which is subsequently used to complete the detection of the action. Experimental results show that this method can ensure high detection accuracy while reducing the computational complexity of the detecting process, so it can achieve the purpose of fast action detection.In order to detect the action area accurately, the selective search algorithm which has achieved effective results in the field of object detection and the Convolution Neural Network are combined to detect actions occurred in video. First, the Convolutional Neural Network are used to filter the right potential action area obtained from selective search. Then greedy search is performed to select a set of subsets where the action occur in the video, and thus to avoid the use of sliding window and improve the accuracy of action detection. Finally, features are extracted from each spatiotemporal subset of action region using the Convolutional Neural Network, and SVMs are used for training and testing. Compared with existing alternatives, this method validates the superior performance of action location as well as competitive results on action detection.Based on the fast action detection algorithm, the action detection software system is designed and implemented. Programmed by Open CV and MFC, the system can detect the specific action in the video automatically and display relevant information in real time with the promotion of multi thread processing strategy.