Efficient Edge Vision

Real-time visual perception within edge power budgets — from FPGA co-design to adaptive detector scheduling

Goal: deliver real-time visual perception on devices where every joule counts, by restructuring the computation — in hardware or at runtime — instead of simply shrinking models.

ALBIREO — adaptive video object detection

A detector-agnostic, training-free framework that wraps off-the-shelf detectors and skips detector invocations when a per-object Kalman state is confident enough; skipped frames get predicted boxes at near-zero GPU cost, with a rescue mechanism preserving objects through brief detector misses. On BDD100K MOT across three detector families and two Jetson generations, accuracy stays within ±1.2 pp of per-frame inference while energy drops 12.1–17.6% — and on the primary configuration accuracy improves by +0.8 pp while energy falls 17.6%. To appear at ACM/IEEE SEC 2026 (Taherin et al., 2026).

Algorithm–architecture co-design for 360° video

Real-time 360° AR/VR video rendering restructured to fit FPGA on-chip memory budgets: co-designing the rendering algorithm with the architecture enabled energy-efficient processing on a Zynq UltraScale+ MPSoC without performance loss versus commercial pipelines. Published at FPGA 2020 (Sun et al., 2020).

References

2026

  1. SEC
    Amir Taherin, José Cano, Bin Ren, and 2 more authors
    In ACM/IEEE Symposium on Edge Computing (SEC), 2026. To appear.

2020

  1. FPGA
    Qiuyue Sun, Amir Taherin, Yawo Siatitse, and 1 more author
    In Proceedings of the 2020 ACM/SIGDA International Symposium on Field-Programmable Gate Arrays (FPGA), 2020