A Tool for Query and Analysis of Mpeg Encoded Video

This thesis presents a tool for efficient query and analysis of MPEG encoded video. The image sequences for the video are obtained via a vehicle-mounted forward-looking camera. The tool accepts queries that signify both temporal and geometric events as the vehicle traverses common roadways, such as " identify portions of the video where the vehicle is making a lane change maneuver, " or " identify portions of the video where the vehicle is going around a tight curve, " etc. Both query and analysis are done directly in the encoded domain. The tool deals with the encoded video stream much like a standard MPEG decoder: § For inter-coded (I) frames of the MPEG video stream, the shape of the lane/pavement markers present in those frames is estimated by a comprehensive search procedure. This lane shape estimation is accomplished in a Bayesian setting using a set of DCT-based lane edge features, a global shape model for lane edges, and a coarse-to-fine optimization algorithm. § The estimation of lane shape in P-frames is somewhat different. Using the motion of macroblocks between frames, and the estimate of lane shape in the previous I-or P-frame, the tool quickly identifies the position and geometry of the lane/pavement markers in the new P-frame. This involves a non-linear least squares fit, followed by a local search in the lane shape parameter space. The lane shape estimation procedure for B-frames is very similar to that for P-frames. § Finally, the lane's geometry in each frame and as it evolves over time is analyzed in order to establish a connection between its signature and the one mandated by the query. Several experimental results are presented to illustrate the efficacy of the tool.