StadiumAI: Real-Time Crowd Density & Flow Analysis with Computer Vision

Stadium Safety & Crowd Dynamics

Large sports stadiums and concert venues face serious crowd stampede risks at entry gates, escalators, and concession zones. In StadiumAI-C4, I developed an intelligent surveillance processing pipeline to monitor pedestrian flow in real time.

Technical Implementation

1. Perspective Transform (Bird’s-Eye View Mapping):

- Converting oblique CCTV camera angles into planar 2D top-down coordinates using homography matrices $H$:

$$\begin{bmatrix} x' \\ y' \\ 1 \end{bmatrix} = H \begin{bmatrix} x \\ y \\ 1 \end{bmatrix}$$

1. YOLOv8 Head & Pedestrian Detection:

- Optimized with TensorRT for $60\text{fps}$ real-time inference on multi-camera streams.

1. Kernel Density Estimation (KDE) Heatmaps:

- Generating 2D Gaussian heat distributions across stadium sectors to highlight high-density congestion zones before dangerous crushes occur.

import cv2
import numpy as np
from ultralytics import YOLO

# Real-time density counting within polygon zones
def calculate_zone_density(frame, model, polygon_coords):
    results = model.track(frame, classes=[0], persist=True, verbose=False)
    boxes = results[0].boxes.xyxy.cpu().numpy()
    
    count_in_zone = 0
    for box in boxes:
        center_x = int((box[0] + box[2]) / 2)
        center_y = int(box[3]) # Feet position
        if cv2.pointPolygonTest(polygon_coords, (center_x, center_y), False) >= 0:
            count_in_zone += 1
            
    return count_in_zone

Practical Use Case

Sends automated alerts to arena control rooms when gate flow capacity exceeds $85\%$ thresholds.