Real-time optical flow processing on embedded GPU: an hardware-aware algorithm to implementation strategy
چکیده
Determining the optical flow of a video is compute-intensive task essential for computer vision. For achieving this processing in real time, whole algorithm deployment chain must be thought efficiency first. The development usually divided into two parts: first, designing an that meets precision constraints, then, implementing and optimizing its execution on targeted platform. We argue unifying those operations enhances performance embedded processor. This paper based industrial use case objective to determine dense time GPU platform: Nvidia AGX Xavier. CLG (combined local–global) method, initially chosen, analyzed understand convergence speed underlying optimization problem. Jacobi solver selected implementation because parallel nature. multi-level then ported GPU, using several specific strategies. In particular, we analyze impact fusing solver’s iterations with roofline model. As result, 30 W power budget, our runs at 60FPS, $$640 \times 512$$ images, four-level processing. Hopefully, example should provide feedback issues arise when trying port method platform serve further implementations vision algorithms specialized hardware.