Learning to perceive objects for autonomous navigation

Current machine perception techniques that typically use segmentation follo wed by object recognition lack the required robustness to cope with the lar ge variety of situations encountered in real-world navigation. Many existin g techniques are brittle in the sense that even minor changes in the expect ed task environment (e.g., different lighting conditions, geometrical disto rtion, etc.) can severely degrade the performance of the system or even mak e it fail completely. In this paper we present a system that achieves robus t performance by using local reinforcement learning to induce a highly adap tive mapping from input images to segmentation strategies for successful re cognition. This is accomplished by using the confidence level of model matc hing as reinforcement to drive learning. Local reinforcement learning gives rises to better improvement in recognition performance. The system is veri fied through experiments on a large set of real images of traffic signs.