Efficient resource allocation for automotive active vision systems

Abstract

Individual mobility on roads has a noticeable impact upon peoples' lives, including traffic accidents resulting in severe, or even lethal injuries. Therefore the main goal when operating a vehicle is to safely participate in road-traffic while minimising the adverse effects on our environment. This goal is pursued by road safety measures ranging from safety-oriented road design to driver assistance systems. The latter require exteroceptive sensors to acquire information about the vehicle's current environment. In this thesis an efficient resource allocation for automotive vision systems is proposed. The notion of allocating resources implies the presence of processes that observe the whole environment and that are able to effeciently direct attentive processes. Directing attention constitutes a decision making process dependent upon the environment it operates in, the goal it pursues, and the sensor resources and computational resources it allocates. The sensor resources considered in this thesis are a subset of the multi-modal sensor system on a test vehicle provided by Audi AG, which is also used to evaluate our proposed resource allocation system. This thesis presents an original contribution in three respects. First, a system architecture designed to efficiently allocate both high-resolution sensor resources and computational expensive processes based upon low-resolution sensor data is proposed. Second, a novel method to estimate 3-D range motion, e cient scan-patterns for spin image based classifiers, and an evaluation of track-to-track fusion algorithms present contributions in the field of data processing methods. Third, a Pareto efficient multi-objective resource allocation method is formalised, implemented, and evaluated using road traffic test sequences.