DCT

Dense crowd tracking

Datasets

5 dataset on human head with bounding box annotations

1. Data_S-HEAD
2. Head Detection (CCTV)
3. Human head detection OpenVM + C270
4. JHU-CROWD++
5. SCUT-HEAD

Training Process

• first attempt with jhu-crowd++ on yolov10
• second attempt with all 5 datasets merged on yolov10

Detectors

• yolo models trained on jhu-crowd++ find the model here at kaggle.
• yolov10 from scratch and pretrained with jhu-crowd++ find the model here at kaggle.

possible modifications

Architecture:

Input:

1024×1024 frame → Split into 256×256 overlapping tiles.

Preprocessing:

FFT High-Pass Filter: Enhances edges of small heads in low-resolution regions.

Circular Hough Attention Layer: Highlights regions with arc-like contours (head candidates).

Backbone:

Hierarchical Dilated Convolution Blocks:

Layer 1: Dilation rate=2 (captures head edges).

Layer 2: Dilation rate=4 (groups nearby heads).

Adaptive Anchors: 3 anchor sizes (4px, 8px, 12px) dynamically assigned via density heatmap.

Head:

Multi-Spectral Classification: Uses RGB + frequency domain features to separate heads from noise.