Implementation of Motion Detection Methods on Embedded Systems: A Performance Comparison

Recently, deploying machine learning methods and deep learning models to create an artificial intelligence system has gained huge interest. Several technologies, such as embedded GPU, ARM multicore processors, visual processor units VPUs, tensor processor units TPUs, and Field programmable arrays FPGAs, have been developed for this purpose. These processors and accelerators have been fitted on different edge boards and single computer boards SBCs. In this work, we present a performance comparison of background subtraction methods with many video resolutions on various technologies and boards. The tested boards are equipped with different versions of ARM multicore processors and embedded GPUs. The aim is to overcome the lack of such studies on embedded devices and compare the performance of these recent hardware configurations. The implementation was achieved on ARM CPUs configuration using OpenCV and on embedded GPU using CUDA OpenCV. Results show that for high computational methods and high-resolution videos, the GPU is four times faster than the CPU. For low-mid computational methods or low-mid video resolution, the GPU performance is reduced due to GPU-CPU bottleneck transfer. This performance comparison enables the reader to better choose the suitable hardware for his mobile application.