Multi-Thread Frame Tiling Model in Concurrent Real-Time Object Detection for Resources Optimization in YOLOv3

Abstract

In the process of object detection, You Look Only Once, version 3 (YOLOv3) approach has been applied as a more efficient solution than Convolutional Neural Network (CNN) methods in accuracy and speed. In some environments, there is a need to process multiple real-time object detection algorithms concurrently where each object detection algorithm receives a live stream from a camera. In this research, we aim to optimize the system resources including the Graphics Processing Unit(GPU) and the Central Processing Unit(CPU) where accuracy indicated using frames per second (FPS) has no change in the process of detection. We propose improvements in terms of architectural models for real-time object detection. In this way, a Multi-thread Frame Tiling model is proposed to optimize GPU and CPU usage per YOLO object to handle concurrent surveillance video streams. Thus, it can be covered a large scale of surveillance cameras on multiple live streams concurrently. To evaluate the proposed model, we investigate the efficiency of GPU and CPU system resources. The experiment results show that the detection algorithm with the Multi-thread Frame Tiling model is by an average of 116.6 in FPS for all video qualities (480p, 720p, and 1080p resolution) compared to the Original YOLOv3 by an average of 27.1 while there is no change in accuracy in the detection process by YOLOv3.