A Visual Navigation Strategy Based on Inverse Perspective Transformation

Vision-Based Navigation Techniques can be roughly divided in map-based and mapless systems. Map-based systems plan routes and their performance and are labeled as deliverative, while mapless systems analyze on-line the environment to determine the route to follow (Bonin et al., 2008). Some reactive vision-based systems include the implementation of local occupancy maps that show the presence of obstacles in the vicinity of the robot and try to perform a symbolic view of the surrounding world. The construction of such maps entails the computation of range and angle of obstacles in a particularly accurate manner. These maps are updated on-line and used to navigate safely (Badal et al., 1994) (Goldberg et al., 2002). Many of the local map-based and visual sonar reactive navigation solutions are vulnerable to the presence of shadows or inter-reflections and they are also vulnerable to textured floors, since they are mostly based on edge computation or on texture segmentation. Solutions based on homography computation fail in scenarios that generate scenes withmultiple planes. Some road line trackers based on Inverse Perspective Transformation (IPT) need to previously find lines in the image that converge to the vanishing point. Some other IPT-based solutions project the whole image onto the ground, increasing the computational cost. This chapter presents a new navigation strategy comprising obstacle detection and avoidance. Unlike previous approaches, the one presented in this chapter avoids back-projecting the whole image, presents a certain robustness to scenarios with textured floors or interreflections, overcomes scenes with multiple different planes and combines a quantitative process with a set of qualitative rules to converge in a robust technique to safely explore unknown environments. The method has been inspired on the visual sonar-based reactive navigation algorithms and implements a new version of the Vector Field Histogram method (Borenstein & Koren, 1991) but here adapted for vision-based systems. The complete algorithm runs in five steps: 1) first, image main features are detected, tracked across consecutive frames, and classified as obstacle or ground using a new algorithm based on IPT; 2) the edge map of the processed frames is computed, and edges comprising obstacle points are discriminated from the rest, emphasizing the obstacle boundaries; 3) range and angle of obstacles located inside a Region of Interest (ROI), centered on the robot and with a fixed radius, are estimated computing the orientation and distance of those obstacle points that are in contact with the floor; 4) a qualitative occupancy map is performed with the data computed in the previous step; and 5) finally, the algorithm computes a vector which steers the robot towards world areas free of obstacles. 5