Use this URL to cite or link to this record in EThOS: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.714863
Title: Moving object detection for automobiles by the shared use of H.264/AVC motion vectors : innovation report
Author: Wong, Chup-Chung
Awarding Body: University of Warwick
Current Institution: University of Warwick
Date of Award: 2015
Availability of Full Text:
Access through EThOS:
Full text unavailable from EThOS. Thesis embargoed until 16 Nov 2018
Access through Institution:
Abstract:
Cost is one of the problems for wider adoption of Advanced Driver Assistance Systems (ADAS) in China. The objective of this research project is to develop a low-cost ADAS by the shared use of motion vectors (MVs) from a H.264/AVC video encoder that was originally designed for video recording only. There were few studies on the use of MVs from video encoders on a moving platform for moving object detection. The main contribution of this research is the novel algorithm proposed to address the problems of moving object detection when MVs from a H.264/AVC encoder are used. It is suitable for mass-produced in-vehicle devices as it combines with MV based moving object detection in order to reduce the cost and complexity of the system, and provides the recording function by default without extra cost. The estimated cost of the proposed system is 50% lower than that making use of the optical flow approach. To reduce the area of region of interest and to account for the real-time computation requirement, a new block based region growth algorithm is used for the road region detection. To account for the small amplitude and limited precision of H.264/AVC MVs on relatively slow moving objects, the detection task separates the region of interest into relatively fast and relatively slow speed regions by examining the amplitude of MVs, the position of focus of expansion and the result of road region detection. Relatively slow moving objects are detected and tracked by the use of generic horizontal and vertical contours of rear-view vehicles. This method has addressed the problem of H.264/AVC encoders that possess limited precision and erroneous motion vectors for relatively slow moving objects and regions near the focus of expansion. Relatively fast moving objects are detected by a two-stage approach. It includes a Hypothesis Generation (HG) and a Hypothesis Verification (HV) stage. This approach addresses the problem that the H.264/AVC MVs are generated for coding efficiency rather than for minimising motion error of objects. The HG stage will report a potential moving object based on clustering the planar parallax residuals satisfying the constraints set out in the algorithm. The HV will verify the existence of the moving object based on the temporal consistency of its displacement in successive frames. The test results show that the vehicle detection rate higher than 90% which is on a par to methods proposed by other authors, and the computation cost is low enough to achieve the real-time performance requirement. An invention patent, one international journal paper and two international conference papers have been either published or accepted, showing the originality of the work in this project. One international journal paper is also under preparation.
Supervisor: Not available Sponsor: Innovation and Technology Commission ; Hong Kong
Qualification Name: Thesis (D.Eng.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.714863  DOI: Not available
Keywords: TL Motor vehicles. Aeronautics. Astronautics
Share: