MTDTS04

The problems to be studied under this proposal deal with collision avoidance for Intelligent Vehicles. We propose to research on techniques for a Multi-Target Detection and Tracking System (MTDTS) for CyberCars, following previous work of the team members. The MTDTS is based on six main modules: - Laserscanner data segmentation -- concerning this subject we propose further research on methods to come up with an improved scan segmentation algorithm, namely improving raw data filtering and adding other line fitting capabilities; - Color camera based object detection and classification --the algorithm will be made up of two submodules:segmentation module using color and geometrical features, provided by the vision system, and using clusters of range data provided by the laserscanner processing module; module for generation of bounding boxes (blobs) defining obstacle regions. A set of different blobs will be considered corresponding each one to a possible different object type. The blobs generation will use information such as dimensions, distance to the object, color; - Object classification using range data - The classification process follows a multi-hypothesis approach and uses a voting scheme considering every hypothesis over time. A confidence level is associated to each classified object; - Object classification fusion - investigate algorithms to robustly fuse information provided by the object classification modules of both laser and vision. Baysian-based methods will be evaluted for this purpose; - Information fusion techniques that will allow to obtain accurate detection and localization of vehicles and pedestrians by making use of data obtained both on-board and off-board the vehicles; - Situation-based information processing -- in this task the algorithms for the different modules of the MTDTS will be analysed and adapted to the specific situations which can arise in CyberCars scenarios; - Object tracking and impact-time computation - the object tracking will be performed by means of a kalman filtering applied to each tracked object. Different process models in the Kalman algorithm fitted to the dynamics of the detected/tracked objects will be researched and used. The impact-time computation will use a classical method based on the projection of all possible points of impact (edges of the vehicle or the object) in the direction of the object´s velocity, and for each instant it is assumed a constant object velocity relative to the vehicle.