Multi-Thread Approach to Object Detection Using YOLOv3

Recently, In the field of the object detection process, Convolutional Neural Network (CNN) methods have been applied as more efficient solutions. In the CNN methods, You Look Only Once, version 3 (YOLOv3) more powerful than classic CNN methods such as Fast-RCNN and Faster-RCNN in terms of accuracy and speed. Even though YOLOv3 can obtain more accuracy and speed compared to other CNN approaches, it needs to be used in a system with a powerful single Graphics Processing Unit (GPU). The YOLOv3 still requires the heavy computational cost to maintain good detection performance; it brings high computation overhead. It is needed to design a more stable algorithm to increase frames per second (fps) in the object detection process. The main contribution is proposing a Multi-thread approach that uses YOLOv3 to perform realtime object detection in the large scale of video streams. In this paper, we design a Multi-thread approach that uses YOLOv3 to perform real-time object detection with decreasing detection time per frame. In the proposed multi-thread approach, we aim to decrease the idle times between CPU and GPU. The proposed approach is evaluated using a public dataset and the result shows improvements in performance compared to the YOLOv3. The multi-thread model is evaluated on two different servers with two GPU cards and the results yield improvement by an average 26% in fps. Contribution-The main contribution is a Multi-thread approach that uses YOLOv3 to perform real-time object detection in the large scale of video streams.