Crowd Counting using CSRNet

Crowd estimation matters for public safety, event management and transport planning, yet manual counting in dense scenes is unreliable — and detection-based methods fail once heads overlap and occlude each other. Drawing a bounding box around every person works in sparse scenes, but in a congested gathering the standard approach collapses exactly where it is needed most. Density-map regression takes a different route: instead of detecting individuals, the network learns "how many people per region," predicting a density map whose pixel sum equals the head count. This degrades gracefully in dense scenes and has been the standard approach since 2016. CSRNet pairs a VGG-16 front-end with dilated convolutions that enlarge the receptive field without downsampling, preserving the spatial detail dense scenes need. For students, it is a clean regression-based vision project with an intuitive demo — upload a photo, get a count plus a heatmap that shows exactly where the model thinks the crowd is.

Dataset & model:
Dataset name: ShanghaiTech Crowd Counting.
Source: publicly released crowd-counting benchmark.
Task: crowd counting via density-map regression.
Classes: not a classification task — head-point annotations converted to Gaussian density maps; Part A: 482 congested images; Part B: 716 street scenes.
Model: CSRNet — VGG-16 front-end (first 10 layers) plus dilated-convolution back-end; trained with Euclidean (L2) loss between predicted and ground-truth density maps.
Input: crowd photograph, resized to a multiple of 8.
Prediction: per-pixel density map.
Output: head count (density-map sum) plus density heatmap overlay.
Evaluation metrics: MAE and MSE — computed by the evaluation notebook on the validation split during the build.
Design target: lower MAE on Part B than on Part A (sparser scenes are easier). This is a design target, not a measured claim; the notebook computes the actual MAE/MSE during the build.

Working:

1. Training phase: head-point annotations are converted into ground-truth density maps by placing a Gaussian kernel at each annotated head position; cropped image patches are augmented at multiple scales for varying head sizes and camera distances.
2. The VGG-16 front-end (first 10 layers) extracts features; the dilated-convolution back-end enlarges the receptive field without downsampling, preserving the spatial detail dense scenes need.
3. Training regresses the predicted density map against the ground truth with a Euclidean (L2) loss on cropped patches; the notebook logs loss curves during the build.
4. Evaluation phase: the evaluation notebook computes MAE and MSE on the validation split and visualizes the error cases.
5. Inference phase: an uploaded crowd photo is resized to a multiple of 8 and passed through the trained network in a single forward pass.
6. The predicted density map is summed to produce the head count, and overlaid as a heatmap on the original photo in the demo app so the count is visually verifiable.

Specifications:
Model | CSRNet: VGG-16 front-end (first 10 layers) + dilated-convolution back-end
Dataset | ShanghaiTech Crowd Counting — Part A (482 congested images), Part B (716 street scenes)
Annotations | Head-point annotations converted to Gaussian density maps
Loss | Euclidean (L2) distance between predicted and ground-truth density maps
Evaluation | MAE and MSE computed by the evaluation notebook on the validation split
Design target | Lower MAE on Part B than Part A (design target, not a measured claim)
Inference | Single-image density-map regression on CPU or GPU
Platform | Windows/Linux/macOS, Python 3.10; web demo app

[Density-Map Head Counting] (implemented) — Counts people from a single crowd photograph by regressing a density map and summing it, with no per-person bounding boxes needed.
[CSRNet Architecture] (implemented) — VGG-16 front-end (first 10 layers) plus a dilated-convolution back-end that preserves spatial resolution while capturing broad crowd context.
[Density Heatmap Overlay] (implemented) — The demo overlays the predicted density map as a heatmap on the photo, showing exactly where the crowd concentrates.
[ShanghaiTech Training] (implemented) — Trains on both Part A (dense crowds) and Part B (sparse street scenes), so the model sees the full range of crowd densities.
[Annotation-to-Density Scripts] (implemented) — Preprocessing scripts convert head-point annotations into Gaussian ground-truth density maps.
[MAE & MSE Evaluation] (implemented) — The evaluation notebook computes Mean Absolute Error and Mean Squared Error on the validation split with error-case visualization.
[Multi-Scale Augmentation] (implemented) — Patch cropping at multiple scales during training handles varying head sizes and camera distances.
[CPU Demo] (implemented) — The shipped weights infer in seconds on a regular laptop CPU.

Complete source code (training, evaluation, inference, demo app)
Trained model weights (PyTorch .pt checkpoints)
Density-map generation and augmentation scripts
Project report PDF (crowd-counting background, methodology, evaluation)
PPT presentation for the final review
Viva Q&A preparation document (density-map regression, dilated convolutions, MAE/MSE, VGG-16)
Setup guide (environment, dependencies, dataset download steps)

Extremely dense crowds saturate the density map — heads packed tighter than the annotation resolution get undercounted.
Occlusion and perspective distortion cause errors: distant small heads and partially hidden people are the main failure modes.
Single-frame estimator: no tracking, so it cannot distinguish individuals or count people across video frames (video counting is documented as future scope).
Educational prototype, not a surveillance-certified system — not intended for safety-critical crowd management or law-enforcement use.
All reported figures are design targets from the buyer's own training run; no pre-measured MAE/MSE is claimed.