FPGA-Based Energy-Efficient Computing Architecture And System Design For Binocular Depth Perception

Binocular stereo vision technology uses two imaging devices in particular locations to obtain images of the same scene and extract three-dimensional information of the scene from the images. It is widely used in reconstruction, virtual reality, augmented reality, automatic drive and other frontier fields. Currently, most binocular stereo vision systems are all based on software platforms. Although with good accuracy, they can not meet the requirements of real-time processing.Therefore a binocular stereo vision system based on FPGA named Smart-Eyes/Depth Perception is proposed. An energy-efficient computing architecture for adaptive window algorithm is designed. And this algorithm is implemented based on the energy-efficient computing architecture using hardware description language. Specific work is as follows:1) Smart-Eyes/Depth Perception system can perform the real-time video acquisition, local-storage, local-processing and transmit the original or processed data to the receiver. It uses two parallel CMOS image sensor to acquire images of maximum resolution of 640*480 and the maximum frame rate of 30fps. The system can transmit and display image data through HDMI interface and can use switch to control the content such as left and right original image or processed disparity image. This paper uses the processing module based on adaptive-weight algorithm to verify the effectiveness of the system function. The system power is about 4.089W and the hardware resources occupancy ratio is about 93%.2) In view of adaptive window algorithm combined with hardware features an energy-efficient computing architecture is designed. According to the architecture the hardware implementation of this algorithm is proposed. Its support window size is 25*25 and maximum disparity level is 64. The functional simulation of this module based on the standard dataset of Middlebury is processed and the rate can reach 30fps. The average error rate is around 13.06%, and the hardware resources occupancy ratio is about 50.74%.In this paper, a binocular stereo vision system based on FPGA called Smart-Eyes/Depth Perception is presented. And on the premise of real-time processing, adaptive window binocular stereo vision algorithm is implemented using hardware description language with good accuracy, large detection range and disparity level. The design make full use of the hardware resources in a reasonable scope and better balance the matching accuracy, system hardware resources and processing speed.