A 154 FPS Efficient Single-frame Bundle Adjustment Accelerator for High Frame Rate Mobile Robots

This paper enhances the efficiency of Visual Odometry (VO) for mobile robots and vehicles by implementing single-frame Bundle Adjustment (BA) on Field-Programmable Gate Arrays (FPGAs). A multi-level reuse hardware architecture is proposed that optimizes the update of the Jacobian matrix and equation solving and helps in reducing FPGA hardware resources and power. Using a parallelizable equation solver, our approach achieves a processing speed of 154 frames per second—2.65 times faster than high-performance CPUs and reduces energy consumption by 8.52 times. Tested on the KITTI dataset, the proposed system reported an average translational error of $1.1 \%$ and a rotational error of $0.0041 \text{deg} / \text{m}$. The results demonstrate FPGAs’, potential to improve VO systems in complex environments.