An Empirical Study of Safety Models used in Driver Assistance Systems from a Statistical Perspective

Driving assistance systems (DAS) are responsible for enhancing the on-road driver experience in terms of lane safety, speed monitoring/control, drowsiness detection, vehicle to vehicle communication for incident alerts, etc. Each of these systems requires large amounts of input temporal data that is processed via machine learning algorithms. For instance, lane safety systems use a combination of image and depth data to analyze whether vehicles are following lane-rules or not. For this analysis, algorithms like convolutional neural networks (CNN), support vector machines (SVMs), and etc. is used. Each of these input-to-algorithm combinations has different advantages and nuances for each application. Thus, it is very difficult for system designers to identify best practices to evaluate and select these algorithms. In order to simplify selection of these algorithms for given systems, this text evaluates different recently proposed and highly efficient systems for each of these applications. This will assist researchers and system designers to select application-algorithm pairs for deploying highly efficient and customized driving assistance systems. The text also suggests various optimizations that can be done in these algorithms to further improve their performance when deployed in new or existing real-time systems.