Real-time crowd density estimation using CSRNet (VGG-16 + dilated CNN) on ShanghaiTech Part A & B. Includes adaptive Gaussian kernel density map generation, from-scratch training on M1 CPU, MPS inference, and a live video demo with heatmap overlay and running count graph. Part A MAE: 87.85 | Part B MAE: 21.12
python computer-vision deep-learning surveillance heatmap pytorch vgg16 density-map smart-city gaussian-kernel video-analytics shanghaitech crowd-counting csrnet dilated-convolutions crowd-density-estimation
-
Updated
Apr 17, 2026 - Python