Research And Design Of ORB Feature Extraction Acceleration Architecture For SLAM Applications

Simultaneous localization and mapping(SLAM)is an important method for robot systems to locate themselves,navigate and construct maps in unknown environments.Visual SLAM(VSLAM)based on the feature matching is an important branch of the SLAM technology.By extracting and matching keypoints in continuous images,the algorithm can predict scene changes and further obtain the pose estimation.Because of the large amount of computation,the algorithm has some defects such as lack of realtime computing and low energy efficiency in traditional embedded applications.Aiming at the computing-intensive feature extraction algorithm in VSLAM,this thesis studies and designs an energy-efficient image feature extraction accelerator for embedded application scenarios,which provides architecture optimization for the application of SLAM technology in energy-efficient embedded scenarios.The main research contents and innovations of this thesis are as follows:1.The performance and computational complexity of existing feature extraction algorithms are analyzed.According to the high real-time requirements of the SLAM applications,the ORB feature extraction algorithm is selected as the accelerated object in this thesis.2.According to the characteristics of the ORB feature extraction algorithm,a streaming acceleration architecture is designed.By reusing the image data and compressing the image data used by descriptor generator,the resource consumption of the row buffers is effectively reduced.3.Aiming at accelerating the most complex part in the ORB feature extraction – the r BRIEF feature descriptor generation,a bucket-based streaming architecture is proposed.By approaching bucket-based keypoints suppression,the detection of unnecessary keypoints and the activation of computing units are reduced.At the same time,each computing unit is redesigned based on the bucket-based dataflow to reduce the resource usage.4.Experiments are designed to evaluate the changed ORB algorithm and the implementation of the accelerator.The experiments show that the changed ORB algorithm can maintain good tracking performance in SLAM applications.Using the TSMC 28 nm process,the accelerator can process 1920 × 1080 full HD images at the frame rate of 120.6FPS at 250 MHz with 73.2m W power consumption.