Computer vision is one of many sensing modalities suited for obstacle detection onboard autonomous vehicles navigating in unknown or only partially known environments. For rotorcraft flying at very low altitudes, the use of image sequences from cameras mounted to the vehicle is desirable because vision is passive and provides rapid sensing of the environment with near infinite range and a wide field of view. Unfortunately, the processing of visual images to generate range maps of the nearby environment is computationally expensive, and currently cannot be performed in real time, even using relatively low-performance cameras. Furthermore, this problem will be exacerbated by the use of new high-performance cameras whose high resolution is necessary to provide the range map denseness required for safe navigation.; We propose a hierarchical set of computations that guarantees that range maps based on feature points can be computed in real time. This is achieved by selecting, at regular intervals in time, a subset of features to track, such that the onboard computing resources are not exceeded. This is enabled and enhanced by an efficient multirate, multiresolutional approach to tracking features. A high level computation determines which features to track, then assigns each feature to one of many parallel slave processors, allocating a maximum amount of computing resources for the feature. Features are chosen to improve overall range map accuracy and feature density, as required for the navigation objectives. Computations at the feature level select image data from a multirate and multiresolution image data structure to produce the best range estimates while using only the available computing resources. The result is a more efficient and more focused processing of image data, more control over the distribution of features in the range map, and a guarantee of real time computing performance.
展开▼