Bucket-FEM: A Bucket-based Architecture of Real-time ORB Feature Extraction and Matching for Embedded SLAM Applications

ORB feature (Oriented-FAST and Rotated-BRIEF) is widely used in SLAM (Simultaneous Localization and Mapping) systems. However, the processing power of embedded platforms is insufficient for the computation-intensive tasks of ORB FEM (feature extraction and matching). Based on our previous work, we propose Bucket-FEM: a bucket-based architecture for real-time ORB FEM accelerators. Firstly, the image is divided into non-overlapping blocks — buckets, and the FEM workflow is organized in the form of buckets. Bucket-based keypoint suppression is applied to reduce power consumption and computation complexity for FEM. Secondly, A novel bucket-based projection matching (BPM) scheme is proposed, which has time complexity of O(n) and delivers high matching performance with minimal resource usage. Finally, the motion-constant assumption is made to predict the model for BPM. The implementation on Vertex-7 XC7V2000T shows that Bucket-FEM supports ORB FEM for FHD images at 68.8fps when running at 142.8MHz, while consuming only 507mW dynamic power and the maximum absolute trajectory error increment is only 0.80mm, which is negligible for most SLAM scenarios.