Image Sequence Processing For Motion Detection And Estimation In Unstructured Environments

Abstract

In real-world decision and control applications requiring motion information as sensory input, the detection and estimation of motion from image sequences is constrained by the available computational resources (time,memory,cpu). In this context, a flexible image processing algorithm which can balance computational resources against desired accuracy in motion estimation has been developed. This algorithm extends the Multistage Hypothesis Testing algorithm [2] to detect and track moving objects in a computationally constrained environment. Unlike standard feature-based or gradient-based techniques for motion analysis, the extended MHT algorithm (1) integrates its motion estimates across multiple image frames, and (2) directs computational resources to the moving regions of interest. The MHT algorithm can be used to detect and track a variety of point and/or linear object features. The initial detection of object features is posed as a rate-constrained detection problem [l]. This ensures that the number of detected features does not overload the computational resources of the motion estimator. The MHT algorithm exploits the timeoptimality of sequential hypothesis testing (Truncated Sequential Probability Ratio Tests) to rapidly search the large space of possible feature trajectories. The resultant detected trajectories are clustered by the hypothesis-tree data structure of the MHT algorithm. This allows for robust motion estimation incorporating information from previous image frames. Multiple trajectory segments are then extended in time, yielding a set of feature trajectories in the image plane. Three-dimensional motion information is then extracted from these 2dimensional feature trajectories. The resulting system provides flexible and efficient motion detection and feature correspondence, within computational constraints, suitable for input to the motion estimation module of an intelligent autonomous system.