Advanced stereo vision disparity calculation and obstacle analysis for intelligent vehicles

• Main Author: Zhang, Zhen
• Other Authors: Dr. Naim Dahnoun ; Prof. Nishan Canagarajah
• Published: University of Bristol 2013
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.629007

Disparity generated from stereo vision based systems provide Driver Assistance Systems (DAS) and Autonomous Transportation Systems (ATS) valuable information to understand the surrounding environment. The research focus is to develop computationally efficient stereo vision based algorithms for disparity calculations and obstacle detections in highway and urban environments. In this thesis, first of all, a novel line based stereo vision disparity calculation is introduced. A high resolution disparity map with a large number of disparity levels is generated using a combination of the pixel's cost function and the disparity of its belonging segment. Secondly, a local cost aggregation algorithm is presented for accurate disparity calculation. A bilateral filter is applied to enhance the normalised cost volume. After quadratic polynomial interpolation, a dense disparity map with sub-pixel resolution is generated. It achieves better accuracy, compared with results from two similar algorithms. Then, an efficient algorithm for on-road obstacle detection is presented. Utilising the disparity characteristic of the road, disparity calculation with reduced search range is performed and the obstacle area is then selected using the error map generated by the matching cost. Furthermore, a novel algorithm for on-road obstacle detection based on stereo cameras is presented. Disparity calculation with full search range is only applied to the bottom row of the image by assuming the initial area of the image is the road surface. The search range of the upper rows can be reduced using the road disparities of the previous line. Hence, road disparities and the obstacle candidates are selected during this process. After that, an extended search range is further applied to obstacle candidates to calculate their disparities. Compared with traditional block-based disparity calculations, it only consumes less than one fifth of the computational power. Finally, a fast local disparity calculation algorithm for automotive applications is introduced. Under the ground obstacle assumption of a typical road scene, only a small fraction of disparity space is required to be visited in order to obtain a disparity map. Information provided by each pixel is utilised and a denser disparity output with low errors in homogeneous areas is generated. Significant amount of improvement in speed and accuracy is achieved while comparing with normal exhaustive search algorithms.