Heterogeneous stereo: A human vision inspired method for general robotics sensing

In this paper, we introduce a heterogeneous stereo sensing system that aim to solve the real life difficulties in vision based autonomous navigation. The common practice for vision autonomous robotics navigation is to concatenate camera(s), SLAM and path planning modules. We observe an dilemma where visual SLAM demand wide angle cameras to perform robust tracking and path planning require sensor to sense long range to perform better. The dilemma exists to both monocular and stereo based system. Such dilemma made the vision only robot can only travel at slow speed or risk crashes. We also take note that current monocular SLAM can do everything that stereo SLAM can do which leads to inefficient uses of the secondary frames. To handle this dilemma and inefficiency, out system rely on two camera with different lenses to achieve maximum information gain with given time frames. Our system fuse observation from consecutive intra frames and spatial inter frames. Intra frames are used to estimate both motion and obstacles at various distance. Spatial inter frames are used to estimate a popper scale whenever possible. To propagate the scale to frames which a good scale can not be observed, monocular ORBSLAM is modified to accept our scale modifier. Using our proposed system, we are able to detect both long range and short range obstacles without using super resolution images or wide baseline stereo rig. We show that by using our proposed system accident such as Tesla autopilot crashes can be avoided.