Rapid precedent-aware pedestrian and car classification on constrained IoT platforms

Demand for computer vision analytics in the embedded world has increased rapidly as the Internet of Things (IoT) expands into cities, workplaces, and homes. Common computationally intensive video and scene analysis tasks, such as pedestrian detection, counting, and tracking, are often relegated to acceleration hardware, or embedded GPUs. This paper showcases decision-making heuristics designed to improve the performance of these analytics. Working within the constraints of low power IoT infrastructure typically utilized in urban, traffic-heavy environments, our Precedent-Aware Classification (PAC) framework provides efficient pedestrian and vehicle detection in the absence of dedicated acceleration hardware. Our implementation takes advantage of frequently traveled routes in order to reduce the amount of required computation, which helps meet the tight timing requirements of embedded platforms where traditional computation models tend to fail. Testing and performance analysis of PAC was done using an ARM Cortex-A9 embedded processor, residing within the Xilinx Zynq 7000 FPGA. In normally populated traffic situations, PAC produced an average 3.23x speed-up and an average 16% improvement in pedestrian detection accuracy over using traditional classifiers alone.