Performance Evaluation of Selected Parallel Object Detection and Tracking Algorithms on an Embedded GPU Platform

Abstract

Performance evaluation of selected complex video processing algorithms, implemented on a parallel, embedded GPU platform Tegra X1, is presented. Three algorithms were chosen for evaluation: a GMM-based object detection algorithm, a particle filter tracking algorithm and an optical flow based algorithm devoted to people counting in a crowd flow. The choice of these algorithms was based on their computational complexity and parallel structure. The aim of the experiments was to assess whether the current generation of low-power, mobile GPUs has sufficient power for running live analysis of video surveillance streams, e.g. in smart cameras, while maintaining energy consumption at a reasonable level. Tests were performed with both a synthetic benchmark and a real video surveillance recording. It was found that the computational power of the tested platform is sufficient for running operations such as background subtraction, but in case of more complex algorithms, such as tracking with particle filters, performance is not satisfactory because of inefficient memory architecture which stalls the processing.