Important Events Detection In Chronical Video Recorded By A Wearable Camera To Monitor Human's Daily Life

As the pervasive application of wearable devices to monitor human's life activities, enormous coarse data are obtained, which guaranteed sufficient samples for medical research and analysis. However, it brings headache troubles to organize and retrieval among the database. In this thesis, we focus on chronical videos recorded by a wearable camera to monitoring calorie intake and consume in human's daily life. Driven by this purpose, approaches to analyze this "abnormal video" based on content are proposed as following:1. Video structuredVideo structured is always the first and a key step in many types of video analysis. According to video recorded by our wearable camera, there is no "shot" in the structure. It is impossible to segment our video by "shots". As the motion of camera is coincident with human, a segmentation method based on global motion is proposed. Firstly, global motions are estimated between frames. After applying DFT to global motion values of frames in each detecting window, two detecting and scalable windows are sliding to locate discontinuity values in a relative larger and size-fixed window. Videos are segmented at discontinuity values. Then, frames with biggest dissimilarity in color histogram are selected in each video segment as key frames to represent this segment.2. Physical activities recognitionUnlike standard activity recognition schemes, the video data captured by our device do not include the wearer him/herself. The physical activity of the wearer is analyzed indirectly through the camera motion extracted from the acquired video. We designed an activity pattern feature based on global motion among a video called Global Motion of Video (GMoV) and trained SVM classifiers by it to indentify simple activities including walk, run, upstairs and downstairs.3. Eating events detection Eating event is important to estimate calorie intake in our application. An automatic detector that finds circular dining plates in key frame is reported for the study of food intake. We first detect contours from input frames. After a number of processing steps that convert contours into arcs, arc filtering and grouping algorithms are applied. Then, convex hulls are identified and the ones that fit the description of ellipses corresponding to dining plates are determined. A simple image mathematical model is built to confirm the efficiency of our algorithm. Our experiments using real-world images indicate that this detector is highly reliable and robust even when the input images contain complex background scenes and the dining plates are severely occluded.Oriented by the application, a framework to analyze monitoring video recorded by wearable camera based on contend is studied, and it is a useful supplement to the whole theory system of Contend Based Video Retrieval (CBVR).