Biologically inspired visual intelligence for unmanned ground vehicles
Date of Issue2012
School of Electrical and Electronic Engineering
Human drivers effectively navigate through most outdoor unstructured environments purely based on their visual inputs. The human visual system has a capability to learn and adapt to its surrounding environment robustly despite the complexity possessed by ground cover variations, uncontrolled lighting, weather conditions, and shadows. However, none of the artificial perceptual systems in the current state of autonomous Unmanned Ground Vehicles (UGVs) have this level of intelligent perception that can continuously cope with the various unexpected situations and perform robustly in dynamic environments. Therefore, understanding the mechanisms underlying the intelligent abilities of a human driver can be the key to design fully autonomous UGVs. The main philosophy underlying this thesis is to understand how humans use their vision systems (eyes) and processing units (brains) in driving scenarios in order to develop humanlike visual intelligence for UGVs that can navigate autonomously in an unstructured outdoor environment. To achieve this goal two important aspects need to be investigated. First aspect of this research addresses the case where a UGV is expected to navigate in a familiar outdoor unstructured environment. Based on empirical evidences, it is known that when human drivers become familiar with the road situations, they prefer to look at the far field where the road edges converge (called the vanishing point) to anticipate the upcoming road trajectory and the car steering with maximal lead time. Based on these findings, vanishing point is considered as a salient and consistent feature during most of the open driving behavior tasks regardless of type of the environments. Hence, in this work, a computational model for the vanishing point estimation based on the visual gaze behavior of drivers is developed for UGVs. The proposed model uses Local Dominant Orientation (LDO) patterns such as road edges, ruts and tire tracks left by previous vehicles to vote for the global convergence point of the road in the visual field. For robust processing of Local Dominant Orientation (LDO), a novel biologically inspired mechanism is proposed utilizing distributed population coding and the opponency mechanism, found in the activities of ensemble of correlated neurons in the primary visual cortex of the mammalian brain. Apart from being biologically plausible, the proposed LDO model significantly outperforms the state of the art LDO methods in terms of both accuracy and robustness in natural and synthetic images.
DRNTU::Engineering::Electrical and electronic engineering::Control and instrumentation::Robotics