A Generic Methodology for Classification of Complex Data Structures in Automotive Industry

Abstract

Driving a vehicle is a complex mixture of real-time processes of cognition and control. Recent advances in pattern recognition and machine learning brought automotive industry to the verge of direct AI applications in vehicles. They would continuously sense the external environment, monitor vehicle’s internal systems and track drivers actions in an attempt to support driver’s decisions, making them better informed or even take over decision process if the reliability and confidence of perceived safety critical situations outperform human performance. This work intends to contribute to the automated real-time classification methodology suitable for applications in the realistic circumstances of driving a vehicle. It proposes a coherent strategy to a fast extraction of simple but robust features out of complex data structures like images, continuous and discrete signals etc. It advocates the use genetic algorithm for feature selection paired with simple classifiers suitable for further combination at the decision level. The presented generic methodology is open for various data transformations, classifiers, features and scales well with the data size. It has been tested in two independent competitions: NISIS Competition 2007 concerned with automated classification of pedestrian images and Ford Classification Challenge 2008 dedicated to symptoms detection from high-frequency signal patterns. The model was announced the winner of the NISIS’2007 Competition achieving pedestrian recognition rate exceeding 95% and is now under evaluation for the second challenge.