|M.Sc Student||Barak Pinkovich|
|Subject||Obstacle Avoidance In an Unstructured Environment|
|Department||Department of Electrical Engineering||Supervisor||Dr. Rotstein Hector|
|Full Thesis text|
This Thesis presents a novel approach to obstacle detection and avoidance that combines two well-known algorithms, one from computer vision and one from control theory. Works on Model Predictive Control for obstacle avoidance usually assume that obstacles to be avoided are computed by an unrelated algorithm. This "black-box" assumption is in principle attractive from a user view-point, but it also suggests the main research question: is it possible to "estimate" the obstacle as part of the state-vector required by Model Predictive Control? The problem in its generality is complex, and therefore this work assumes relatively simple sensing: the only information available from the environment is a monocular camera. The sequence of image frames acquired from the camera is considered the measurement process, so that the algorithm used for detecting obstacles can be seen as an estimator, estimating both navigation errors and a meaningful description of the environment. The algorithm of choice is the Bundle Adjustment, which is an optimal non-linear quadratic estimator and hence blends nicely with the Model Predictive Control quadratic penalty function. The Bundle Adjustment estimates the location of features that can be tracked from image to image on the three-dimensional world. A straightforward approach would then be to implement some sort of segmentation to extract potential obstacles from the background. This approach has two main difficulties: first, segmentation is by itself a computational expensive procedure, and second it needs to be repeated continuously since obstacles are "discovered" as they are completed when new views become available. Indeed, three dimensional appearance cannot be inferred from partial views in an unstructured environment. To overcome these difficulties, obstacles were modeled directly using point clouds, which are three dimensional features with location computed in the vehicle’s frame. In a central result, it will be shown that an existing Model Predictive Control algorithm, which has been successfully used and tested, can be adapted to cope with the new way of describing obstacle. Based on this description, a framework is specified that combines a quadratic estimator with a quadratic control scheme. Although no overall optimality is claimed or even pursued, the fact that similar criteria are used helps motivate the overall approach.
Significant effort will be placed in proving the feasibility of the control scheme. First, a standard data-base called KITTI was processed with the objective of implementing an off-line simulation. Then, a scale scenario was implemented to do actual testing. This scenario includes a scaled-down vehicle provided with a monocular camera, a "world scene" that allows generating images and an external source of navigation information. The simulation stage was not successful and going directly from problem formulation/implementation to experimental work is in general not recommended, but the testing showed that the approach is worth pursuing.