Approaching Camera-based Real-World Navigation Using Object Recognition

Traditional autonomous navigation systems for transportation use laser range scanners to con- struct 3D driving scenes in terms of open and occupied voxels. Active laser range scanners suffer from a series of failures, such as inability to detect wet road surfaces, dark surfaces and objects at large distances. In contrast, passive video cameras are immune from these failures but processing is challenging. High dimensionality of the input image requires efficient Big Data analytic methods for the system to perform in real-time. In this paper we argue that object recognition is essential for a navigation system to generalize learned landmarks to new driving scenes, which is a requirement for practical driving. To overcome this difficulty we present an online learning neural network for indoor navigation using only stereo cameras. The network can learn a Finite Automaton (FA) for the driving problem. Transition of the FA depends on several information sources: sensory input (stereo camera images) and motor input (i.e. object, action, GPS, and attention). Our agent simulates the transition of the FA by developing internal representation using the Developmental Network (DN) without handcrafting states or transi- tion rules. Although the proposed network is meant for both indoor and outdoor navigation, it has been only tested in indoor environments in current work. Our experiments demonstrate the agent learned to recognize landmarks and the corresponding actions (e.g. follow the GPS input, correct current direction, and avoid obstacles). Our future work includes training and learning in outdoor driving scenarios.