Tracking and Classification of Dynamic Obstacles Using Laser Range Finder and Vision

A multi-module architecture to detect, track and classify objects in semi-structured outdoor scenarios for intelligent vehicles is proposed in this paper. In order to fulfill this task it was used the information provided by a laser range finder (LRF) and a monocular camera. The detection and tracking phases are performed in the LRF space, and the object classification methods work both in laser (with a Majority Voting scheme and a Gaussian Mixture Model (GMM) classifier) and in vision spaces (AdaBoost classifier). A sum decision rule based on the Bayes approach is used in order to combine the results of each classification technique, and hence a more reliable object classification is achieved. Experiments using real data confirm the robustness of the proposed architecture.