Virtual Adaptive and Real-Time Monocular Car Vision and Navigation Using Modular Arti cial Neural Networks In

Abstract

The perceived complexity of this problem has led to generation of complex solutions using complex multi-sensor control architectures to realize the so called intelligent cars. While complex multi-sensor control architectures o er great performance, their operational speeds are still low for realtime high speed driving applications. In this thesis, speed optimized algorithm for Advanced Driver Assistance System (ADAS) that operates realtime on high speed driving is developed. Virtual navigation ar- chitecture in which a virtual car, car's kinematic model, linked to the physical car's odometry self-drive on a virtual map derived from monocular vision is proposed. Due to the heterogeneity of the driving environment, a multi model approach and comple- mentary adaptive, fast model-switching scheme, is developed using hidden Markov models. A lean explicit path planning method using single deformable poly-line with velocity objects and a single robust path tracking controller based on backstepping algorithm are developed.Suitability of arti cial neural networks in this problem is demonstrated. They have been used to clearly segmente road scenes using colour cues while using surface detection as opposed to edge detection, used to designate keypoints operating as edge detector and also used in range computation. This work further demonstrated how speed of processing can be improved through a variety of methods; saccadic vision, sparse keypoints, adaptive model switching, single robust controller, modular neural network, multithreaded algorithm and virtual navigation framework. Above all, this work manifest a lean and fast ADAS that can support realtime high speed driving. XV