Outdoor mobile robots usually rely on visual landmarks for guidance and navigation purposes. However the task of locating these landmarks in cluttered environment is very complex. The perceived size of the landmarks is a function of the imaging distance of the robot. Also, the appearance of the landmark may vary due to changes in lighting, viewing direction and interfering objects. In this paper we present a fractal based approach to landmark detection. Fractals are inherently scale invariant; the model presented here is also invariant to changes in light intensity. We present the performance of the incremental fractional Brownian motion (IFBM) classifier for natural textures as well as a segmentation tool for natural scene images.>

A new approach using fractal based vision is presented to solve the problem of mobile robot navigation in outdoor environments. Mobile robots rely on landmarks such as mile markers and street signs for information on global position and local traffic conditions. Due to the motion of the robot, the location, size and orientation of the landmarks are varying. Also, other objects in the scene might partially occlude the landmark. Thus, a robust recognition system is required to recognize the landmarks that may be distorted by a combination of these effects. A new fractal model called incremental fractional Brownian motion (BM) model, is developed to locate these landmarks. A new neural network architecture, reconfigurable neural network (RNN), is developed to recognize the landmarks. The fractal model is shown to be invariant to changes in intensity of incident light. The landmark candidate regions detected by the ifBM model are analyzed by the RNN. New learning rules based on update normalization are developed to decrease learning time and increase system stability. The network also has the ability to learn new patterns with minimal retraining time. The network is tested with images of actual street signs that were distorted by scale changes, rotations and occlusions.

In this paper we present a neural net- work based approach to landmark recognition. Landmarks are used by mobile robots for navi- gation and self referencing information. Due to the motion of the robot, the location and size of the landmark in the sensor is always changing. The network architecture presented here is capable of overcoming changes in scale as well as position of the landmark. The neu- ral network is directly fed with intensity im- ages. This is a significant reduction in compu- tation overhead since no high-level image pro- cessing is needed. Since robot motion in un- structured environments requires dynamic path planning, the detection algorithm must also be able to output guidance results. The network described here is able to output results in real time for the input patterns. We have trained the network to recognize common landmarks such as road signs (stop sign, yield sign, pedestrian-crossing sign, etc.). The algorithm has been successfully integrated into a mobile robot testbed developed at the Robotics and In- telligent Systems Laboratory at North Carolina State University.