Crowd Counting from Images using CNN

Counting people in dense crowds defeats bounding-box detectors: heads merge, overlap and shrink with perspective until boxes are meaningless. Density-map regression sidesteps detection entirely — the network predicts a per-pixel density whose spatial sum is the head count. The ShanghaiTech dataset (Zhang et al.) provides 1,198 images with 330,165 head-center annotations across dense internet scenes (Part A) and Shanghai street scenes (Part B): the standard benchmark. CSRNet showed that dilated convolutions give counting the large receptive fields it needs without pooling away resolution. This project trains that architecture, evaluates MAE/MSE on the official test splits, and demos heatmap overlays with overcrowding alerts for venue-safety use.

1. Head-center annotations are blurred with Gaussian kernels to create ground-truth density maps.
2. A VGG-16 frontend (first 10 conv layers, ImageNet-initialized) extracts features.
3. Six dilated convolution layers (rate 2) expand the receptive field without downsampling.
4. The network outputs a single-channel density map at 1/8 input resolution, trained with Euclidean loss.
5. Summing the map gives the head count; thresholding it into bands gives zone alerts.
6. Training runs on ShanghaiTech Parts A and B; evaluation reports MAE/MSE per official split.

Tech stack:

• Python 3.10, PyTorch
• CSRNet-style fully-convolutional network
• ShanghaiTech dataset (Zhang et al.)
• OpenCV, NumPy (density maps, overlays)
• Flask (demo backend)
• HTML/CSS/JavaScript (demo UI)
• Git

Dataset & Model Details

• Dataset: ShanghaiTech — 1,198 images, 330,165 head-center annotations; Part A: 482 dense internet scenes; Part B: 716 Shanghai street scenes; source: Y. Zhang et al., ShanghaiTech University.
• Model input: crowd scene image. Output: single-channel density map; count = spatial sum.
• Architecture: VGG-16 frontend (10 conv layers) → 6 dilated conv layers (rate 2) → density map; Euclidean loss vs. Gaussian-blurred ground truth.
• Metric: MAE / MSE of counts on the official test splits, measured after training. No figure is claimed before the training run.

• [Density-map regression] Fully-convolutional network predicts a per-pixel density map; the count is its spatial sum — no boxes needed.
• [CSRNet-style architecture] VGG-16 frontend plus six dilated convolution layers for large receptive fields without resolution loss.
• [Heatmap overlays] Predicted density rendered green→yellow→red over the scene so crowd concentration is visible.
• [Overcrowding alerts] Zone-level density thresholds raise LOW/MEDIUM/HIGH badges against a configurable persons/m² limit.
• [Gaussian ground truth] Head annotations blurred with geometry-adaptive kernels to build training density maps.
• [Honest split evaluation] MAE/MSE measured separately on Part A (dense) and Part B (street) test splits.
• [Scene demo] Upload a crowd photo; get the heatmap, count and alert status.

• Python training notebook (density-map training, Gaussian GT generation, evaluation)
• Density-map inference pipeline + heatmap visualization code
• Trained model weights
• Test-split evaluation: MAE/MSE on Parts A and B from the actual run
• Web demo (scene upload, heatmap overlay, count, alert badges)
• Project report PDF (methodology, split comparison, measured results)
• PPT presentation for final review
• Viva Q&A preparation document