Gait-based human recognition in video

Gait, characterizing the manner of human walking, can be used to recognize non-cooperating humans at a distance. This dissertation addresses several fundamental problems for gait-based human recognition in color and infrared videos under real-world environments.; A new spatio-temporal gait representation, called Gait Energy Image (GEI), is proposed for model-free human recognition. GEI not only saves storage space and computation time but also it is less sensitive to noise in individual frames. To address the problem of large intra-person variation of gait appearance under different environmental conditions and lack of training data, the model-free approach generates real and synthetic gait templates and combines them to improve the human recognition performance. The same model-free framework has been extended to view-insensitive gait recognition, and activity recognition in video.; Bayesian-based statistical analysis is performed to evaluate the discriminating power of static gait features for human recognition. Using the available anthropometric data and probabilistic analysis, the upper bound on probability of correct recognition for different resolutions of human appearance in video is obtained. A model-based approach is then proposed for human recognition by gait. The proposed approach estimates 3D human walking parameters by fitting a 3D kinematic model to a video. This approach is first applied for gait recognition from a single video camera, and then extended to multiple video cameras.; To improve the performance of moving human detection for both model-free and model-based human recognition, the information from color and infrared videos is combined using automatic image registration. A hierarchical scheme is developed to automatically find the correspondence between synchronous color and infrared videos based on the human motion in the scene. Strategies for probabilistically combining cues from registered color and infrared videos for improved human detection are presented.