ARTIGO TÉCNICO M. Oliveira, V. Santos Member, IEEE Department of Mechanical Engineering University of Aveiro, Portugal {mriem, vitor}@ua.pt
MULTI-CAMERA ACTIVE PERCEPTION SYSTEM WITH VARIABLE IMAGE PERSPECTIVE FOR MOBILE ROBOT NAVIGATION ABSTRACT This paper describes a system and associated procedures to efficiently manage and integrate dynamic multi-perspective images for robot navigation on a road-like environment. A multi-camera device atop a pan-and-tilt unit has been conceived for enhanced perception and navigation capabilities. These capabilities will cover for active perception, dynamic tracking, foveated vision and, possibly in the future, stereo perception. Due to the large variation of points of view, the vision unit generates images with very different perspectives distortions which inhibit image usage without geometric correction either for image fusion or geometric evaluation of targets and objects to track. By using a versatile kinematics model and efficient multi-camera remapping techniques, a procedure was developed to reliably detect the road and features. The systems is now capable of generating a bird view of the road from several images with different perspectives, in under 30 ms on a 1.8GHz Dual Core machine, enabling the application of this technique in real time navigation systems.
I. INTRODUCTION In the context of the Portuguese Robotics Open (Festival Nacional de Robótica) competition of Autonomous Driving [1][2], several issues arise when robot navigation is to be performed solely based on vision. The main goal of the contest is to traverse a road-like track seeded with several additional difficulties such as a zebra crossing area, a mid-road dashed line, traffic lights and other limiting demands (Figure 1). Fast and reliable execution of this task using vision requires large amounts of image acquisition and processing both to mind the road and to perceive other information such as signs and indications, and even unknown obstacles. A single fixed camera with common optics makes the task fairly complex due to the wide field to be observed, so the solution often ends up in using two or more cameras to split the perception focus and keep the desired pace of fast motion with safety and respect for signs and other directives.
Figure 1 . Autonomous Driving competition environment.
[8]
robótica
When using limited or fixed cameras, the navigation problem is further increased for some types of maneuvers such as parking on a delimited area; it is clear that when reaching immediately above the delimited area the perception is much more difficult because the target vanishes and no references exist to track. Another example might be what may be called dynamic tracking which occurs when the target (obviously) moves but also the tracking system moves with its own motion law, increasing the complexity of the overall relative motion. An example of this problem is car overcoming on real roads. The tracking problem in wide fields is by itself a demanding issue due to the tradeoff between image detail and speed of processing. Current trends [3][4] are converging into foveated vision where attention mechanisms operate on a low resolution wide field image and tele-lenses, or other
Figure 2 . Atlas MV Robot with the multi-camera perception unit.