Live Video Analytics with FPGA-based Smart Cameras

Analyzing video feeds from large camera networks requires enormous compute and bandwidth. Edge computing has been proposed to ease the burden by bringing resources to the proximity of data. However, the number of cameras keeps growing and the associated computing resources on edge will again fall in short. To fundamentally solve the resource scarcity problem and make edge-based live video analytics scalable, we present an FPGA-based smart camera design that enables efficient in-situ streaming processing to meet the stringent low-power, energy-efficient, low-latency requirements of edge vision applications. By leveraging FPGA's intrinsic properties of architecture efficiency and exploiting its hardware support for parallelism, we demonstrate a 49x speedup over CPU and 6.4x more energy-efficiency than GPU, verified using a background subtraction algorithm.