An Analytical Comparison of Approaches to Real-Time Object Detection to Handle Concurrent Surveillance Video Streams

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. To cover a large scale of surveillance cameras on multiple live streams concurrently, we have to decrease the detection time per frame while performance does not make changes. Since there is no comprehensive study of known approaches to object detection in the camera surveillance environment, we conduct a state-of-the-art study to compare the approaches Single Shot MultiBox Detector (SSD), Region-Convolutional Neural Network (FASTER R-CNN), You Only Look Once version 3(YOLOv3), YOLOv4, and Scaled YOLOv4 in different Test servers. We use the multi-thread model that we proposed in our last work on multiple live streams to improve the efficiency of GPU and CPU system resources. The experiment results show that the Multi-thread model's detection algorithm is approximately 27%, 26%, and 22% faster than the original YOLOv3, YOLOv4, and Scaled YOLOv4 in all Test servers on video resolutions while there is no change inaccuracy in the detection process by detection algorithms. Furthermore, the multi-thread models are evaluated on four different servers with four GPU cards. The results show that the Analysis server RTX 4000 has the best result between provided Test servers.