Abstract
Real-time stereo vision is attractive in many areas such as outdoor mapping and navigation. As a popular accelerator in the image processing field, GPU is widely used for the studies of the stereo vision algorithms. Recently, many stereo vision systems on GPU have achieved low error rate, as a result of the development of deep learning. However, their processing speed is normally far from the real-time requirement. In this paper, we propose a real-time stereo vision system on GPU for the high-resolution images. This system also maintains a low error rate compared with other fast systems. In our approach, the image is resized to reduce the computational complexity and to realize the real-time processing. The low error rate is kept by using the cost aggregation with multiple blocks, secondary matching and sub-pixel estimation. Its processing speed is 41 fps for $2888\times 1920$ pixels images when the maximum disparity is 760.
Highlights
The aim of stereo vision systems is to reconstruct the 3-D geometry of a scene from images taken by two separate cameras
In order to achieve the balance of the processing speed and the accuracy, we focus on proposing a series of methods to accelerate the multi-block matching (MBM) algorithm proposed by [9] on GPU
According to the evaluation results in the KITTI Benchmark 2015, the error rates of Recurrent Neural Network (RNN) [10] (6.34%) and the Semi-Global Matching (SGM) [11] (8.24%) is higher than the MBM [9] (6.04%), which is used in this paper, because both of them need to simplify the algorithms to fit the hardware architecture
Summary
The aim of stereo vision systems is to reconstruct the 3-D geometry of a scene from images taken by two separate cameras. Sophisticated algorithms such as [5] and [6] have been implemented on GPU, and they showed very high accuracy Their processing speed for a high resolution image set is far from the real-time requirement. References [1] and [8] implemented the minimal spanning tree (MST) and dynamic programming (DP) on FPGA respectively They achieved the real-time processing (30fps) for the high resolution images, but all of their disparities are smaller than 64, which is not suitable to the modern requirements. Reference [10] implemented a Recurrent Neural Network (RNN) aggregation method on a highend GPU Geforce GTX Titan X, and [11] implemented a Semi-Global Matching (SGM) method on a low-end embedded GPU Tegra X1 In both of these two GPU systems, the global algorithms are processed in real-time. According to the evaluation results in the KITTI Benchmark 2015, the error rates of RNN [10] (6.34%) and the SGM [11] (8.24%) is higher than the MBM [9] (6.04%), which is used in this paper, because both of them need to simplify the algorithms to fit the hardware architecture
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
