Project Overview

Entry Exit Count AI is a real-time computer vision system designed to automatically detect, track, and count people entering and exiting a monitored area using CCTV or video feeds. The system uses advanced object detection (YOLOv8) and multi-object tracking (DeepSORT) to provide accurate IN and OUT counts.

The Flask-based web dashboard provides live video streaming, real-time analytics, historical data, and crowd safety alerts. Perfect for malls, offices, institutions, and public events requiring automated crowd monitoring.

Problem Statement

Traditional CCTV systems only record footage without providing actionable insights. Manual people counting is:

• Time-consuming and labor-intensive
• Error-prone, especially in crowded environments
• Expensive to scale across multiple locations
• Lacks real-time insights for immediate decision-making

Solution Approach

Implemented an AI-powered computer vision pipeline:

• Detection: YOLOv8 for real-time person detection in video frames
• Tracking: DeepSORT algorithm for multi-object tracking across frames
• Analysis: Virtual line-crossing detection for direction determination
• Dashboard: Flask web app with live streaming and analytics
• Alerts: Intelligent crowd safety notifications

Key Features

Technical Stack

Computer Vision

Backend & Web

Frontend & Visualization

Implementation Details

Video Processing Pipeline

• Capture frames from CCTV/video source
• Detect persons using YOLOv8 model
• Extract bounding boxes and confidence scores
• Track detections across frames using DeepSORT
• Determine if track crosses virtual counting line
• Update counters based on crossing direction

Key Components

• Detection Module: YOLOv8 pre-trained model for person detection
• Tracking Module: DeepSORT for persistent multi-object tracking
• Counting Logic: Virtual line intersection algorithm
• Web Interface: Flask app with WebSocket for live streaming
• Data Persistence: SQLite database for historical records

Line-Crossing Algorithm

• Define virtual line coordinates in video frame
• Track centroid of each detection across frames
• Detect when centroid crosses the virtual line
• Determine direction based on movement vector
• Increment appropriate counter (IN or OUT)

Challenges & Solutions

Challenge 1: Occlusion Handling

Problem: People getting occluded by objects or other people.

Solution: DeepSORT maintains track IDs even during temporary occlusions using appearance features and Kalman filter predictions.

Challenge 2: False Positives

Problem: Detection of non-person objects as people.

Solution: Applied confidence thresholds and post-processing filters to reduce false positives.

Challenge 3: Performance Optimization

Problem: Real-time processing demands on video stream.

Solution: Optimized frame processing, used GPU acceleration, and implemented frame skipping strategies.

Real-world Applications

• Shopping Malls: Track customer traffic and peak hours
• Office Buildings: Monitor occupancy and desk utilization
• Institutions: Track attendance in real-time
• Public Events: Manage crowd flow and capacity
• Airports/Transit: Monitor passenger flow
• Smart Cities: Crowd management and safety

Project Links