Computer vision and deep learning are emerging technologies as the backbone system to maintain the public healthcare sector to detect objects and surroundings, especially during the COVID-19 pandemic. Generally, in a single stage, you only look once version 3 (YOLOv3) algorithms promising the best results to detect the object in images, live feeds, or videos by learning features at a faster rate than two-stage algorithms such as R-CNN, fast CNN, and faster CNN. Deep sort methods were employed to track identified people by supporting bounding boxes and calculating the Euclidian distances between the people to maintain social distance. Moreover, the YOLOV3 model requires more computational cost to detect the object at best with a lower detection time. Hence, it motivates us to practice a single graphics processing unit (GPU) with the multithreaded approach to increase the frames per second at detection. The proposed model uses a background modeling method grounded on frame variance accumulation which is used to define the number of frames and weight updating. This approach uses two steps, localization of the object and then the classification of localized objects. Distances between people are calculated and compared with threshold values to facilitate comparison. The threshold limit triggers the alert system which is accessible to people, monitoring many video streams at a time. The model is tested based on processors, threads consumed, and various types of inputs ranging from static images to moving videos. Tiny-YOLOv3 performs with the best frames per second and the least processing time, followed by SPP-YOLOv3 and YOLOv3. The model proves its evidence on various parameters and metrics to work robustly. As well as the reason for adopting YOLOv3 over other YOLOv4 and YOLOV5 is tabulated. This model initiates the curiosity to develop a mobile application with security systems based on IoT and CCTV to monitor crowded places.