Enhancing Real-Time Human Tracking using YOLONAS-DeepSort Fusion Models

Real-time human detection and tracking are dynamic research domains within the computer vision community. While human detection has made significant advancements in recent years, the challenge of detecting humans in highly crowded environments, especially in the presence of occlusions, remains largely unresolved. Over the years, various methods for real-time human detection and tracking have garnered significant attention due to their extensive applications in crime detection, people counting, public event management, disaster management, safety monitoring, and more. In this proposed research, we employ YOLONAS (You Look Only Once Neural Architecture Structure) for the detection and localization of objects within images or video frames of the benchmark crowd-based datasets. Subsequently, we utilize the DeepSORT (Deep Simple Online and Real-time Tracking) algorithm to track objects across frames in video sequences. The research also involves a comparison of the performance of the proposed approach on different datasets, including COCO-Persons, Football Players Image, and CityPersons. Hence, the combination of YOLONAS and Deepsort is a powerful architecture for detecting and tracking humans and achieved above 90% detection accuracy on all the datasets. Experiments were conducted on the Jetson Nano Developer Kit, showcasing the system's ability to process high-resolution sensors and run multiple neural networks simultaneously. The proposed system demonstrates outstanding performance, surpassing state-of-the-art methods in small object detection, localization accuracy, post-training quantization, and real-time edge-device application.