City Research Online

Abstract

This thesis introduces a novel framework for a real-time occupant detection system capable of extracting both two- and three-dimensional information using a single imager with active illumination. The primary objective of this thesis is to demonstrate the feasibility of such a low-cost classification system with comparable performance to multi-camera based stereo vision systems. Severe illumination conditions characterised by a frequent and wide illumination fluctuation are also challenging problems addressed in this work. The proposed system is designed to solve a problem of classifying three occupant classes being an adult, a forward-facing child seat, and a rear-facing child seat.

DoubleFlash is employed to eliminate the influence of ambient illumination and to compress the optical dynamic range of target scenes. The idea underlying this technique is to subtract images flashed by different illumination power levels. The extension of this active illumination technique leads to the development of a novel shadow removal technique, called ShadowFlash. By simulating an artificial infinite illuminating plane over the field of view, the technique produces a shadowless scene without losing image details by composing multiple images illuminated from different directions. The ShadowFlash technique is then extended to the temporal domain by employing the sliding n-tuple strategy, which is introduced to avoid the reduction of the original frame rate.

A modified active contour model, facilitated by morphological operations, extracts the boundary of the target object from the shadow-free scenes produced by the ShadowFlash. Based on the brightness information of the image triplet generated by the DoubleFlash, the orientations of the object surface at pixel points are estimated by the photometric stereo method and integrated into the 3D surface by means of global minimisation. The boundary information is used to specify the region of interest to reconstruct. Investigating both the two- and three-dimensional properties of vehicle occupants, 29 features are defined for the training of a neural network. The system is tested on a database of over 84,000 frames collected from a wide range of objects in various illumination conditions. A classification accuracy of 98.9% was achieved within the decision-time limit of three seconds.

Publication Type:	Thesis (Doctoral)
Subjects:	T Technology > TA Engineering (General). Civil engineering (General) T Technology > TK Electrical engineering. Electronics Nuclear engineering
Departments:	School of Science & Technology > School of Science & Technology Doctoral Theses Doctoral Theses School of Science & Technology > Engineering

Export

Downloads