AI Vision-based Software for Single Camera Systems: Mapping Human Heat Density

This paper presents creating a human detection and counting software with a heat map density feature that can accept image or video input and output the number of humans detected with bounding boxes. A heat map allows for easy visualization of human density in an area or image, allowing for easier foot traffic monitoring. Multiple convolutional neural network (CNN) models were tested and compared to be used as artificial intelligence (AI) for human detection of images and videos. The research method used is quantitative. It was used to determine the best CNN model between YOLOv5-crowd or YOLOv5m, YOLOv5s, R50-FPN 3x, and SSD. Additionally, it was used to determine the best confidence and IoU threshold to optimize the best CNN model. The datasets used for these methods were composed of thirty randomly selected images of non-overlapping humans, crowded/overlapping humans, and from a chosen video. There were three tests conducted for each method with each dataset. These tests are accuracy, precision, and speed. Once the best CNN model and its best parameter have been determined. This software includes a heat map functionality that is composed of a current heat map and an overall heat map. The results showed that YOLOv5-crowd or YOLOv5m is the best CNN model with its optimized value of 25% confidence threshold and 15% IoU threshold.