Estimation d'états pour le pistage par lidar à faible résolution angulaire
|Abstract:||In this master's thesis, state estimators are designed to track targets seen by a lidar with very low angular resolution. The Hough transform is modified in order to consider the polar configuration of the lidar detections, the gaussian noise on their distance measurements and the uniform noise on their angular position measurements. The designed Hough transform estimates the trajectory of a target by evaluating the best line passing on the most recent detections. This estimation approaches the maximum likelihood. Some methods to reduce the computational complexity are also shown. The results of the Hough transform are then combined with Kalman filters (classic, extended and unscented). Two main strategies are developed : transforming the uniform noise into gaussian noise by projecting the angular position on the estimated line, or adding the line as soft constraints in the observation vector of the Kalman filter. Simulations show that the unscented Kalman filter with soft constraints has the lowest mean square error on the state estimation on average. Moreover, this method's improvement in accuracy, in comparaison to a regular unscented Kalman filter, is significant in the following cases : when the target moves away, when the target is far away or when the target is smaller than the width of a lidar element. Simulations are done for different linear and curved trajectories, for two sizes of targets (pedestrian and car) and for three lidar configurations (16 elements of 2:8° each, 8 elements of 5° each et 9 elements of 10° each). Experimental results for pedestrians and a 16 elements lidar are also discussed.|
|Document Type:||Mémoire de maîtrise|
|Open Access Date:||7 May 2018|
|Collection:||Thèses et mémoires|
All documents in CorpusUL are protected by Copyright Act of Canada.