Traffic Analysis and Vehicle Counting System

A robust, intelligent traffic monitoring system built to detect, track, and analyze vehicle movement in diverse conditions using advanced computer vision and transformer-based AI models.

Team Members:

Mayank Jangid, Kushal Khemka, Abhinav Rajput

---

Project Overview

This system is designed to go beyond basic vehicle detection — offering:

• Real-time vehicle counting
• Multi-object tracking
• Speed estimation
• Automatic Number Plate Recognition (ANPR)
• Congestion detection
• Weather-resilient detection
• Vision-Language scene understanding

---

Key Features

Object Detection & Model Evaluation

Evaluated 8 models across precision, recall, and AP metrics:

Model	AP50	AP50_95	Precision	Recall
RF-DETR-Large	0.985	0.790	0.967	0.957
YOLOv8	0.977	0.779	0.955	0.932
YOLOv12	0.974	0.774	0.922	0.945
YOLOv11	0.974	0.774	0.930	0.947
RT-DETRv3-R50	0.968	0.765	0.942	0.939
RT-DETRv2-R101	0.961	0.759	0.937	0.935
RT-DETRv2-R50	0.957	0.752	0.931	0.930
RT-DETR-X	0.935	0.742	0.907	0.920

Recommendations:

• RF-DETR-Large: Best for precision-critical deployments
• YOLOv8: Best accuracy-efficiency tradeoff
• YOLOv12: For using latest YOLO advancements
• RT-DETRv3-R50: End-to-end transformer solution

---

Weather-Robust Augmentation

To maintain high performance under fog, rain, and snow:

• Analytical Augmentation: Depth-based weather simulation
• Neural Style Transfer: Stylistic weather transformations

Result: Improved generalization under challenging environmental conditions

---

Tracking & Analytics

Multi-Object Tracking with ByteTrack

• High accuracy tracking via low- and high-confidence box associations
• Handles occlusions and low-light scenarios effectively

Dwell Time & Congestion Detection

• Custom ROI tracking for zone-specific monitoring
• EMA-based congestion detection using statistical thresholding
• Supports dynamic adjustment based on location-specific historical data

---

Automatic Number Plate Recognition (ANPR)

• YOLO-based plate detection
• OCR for character extraction
• Works under varying lighting, angles, and plate formats

---

Speed Estimation

• Frame-to-frame pixel displacement analysis
• Supports relative speed calculation with multi-object tracking
• Highlights need for accurate camera calibration for real-world units

---

Vision-Language Integration

Using Qwen-2.5VL-7B:

• Scene understanding and description
• Object relationship detection
• Natural language summaries of traffic flow and congestion events
• Bounding box and point-based object localization

---

System Architecture

[Video Input] → [Object Detection] → [ByteTrack Tracker]
             → [ROI Analyzer] → [Dwell Time, Congestion]
             → [Speed Estimator] 
             → [ANPR Module]
             → [Qwen2.5-VL Scene Description]

---

References

• RF-DETR: https://github.com/roboflow/rf-detr
• Qwen2.5-VL: https://github.com/QwenLM/Qwen-VL
• ByteTrack: https://github.com/ifzhang/ByteTrack
• Weather Augmentation Paper: https://openaccess.thecvf.com/content/WACV2024/papers/Gupta_Robust_Object_Detection_in_Challenging_Weather_Conditions_WACV_2024_paper.pdf
• PaliGemma 2: https://developers.googleblog.com/en/introducing-paligemma-2-mix/

---

Future Work

• Real-world speed enforcement with geometric calibration
• Deployment on embedded edge devices
• Integration with traffic signal optimization systems

---

Clone the repository

git clone https://github.com/mayank-jangid-moon/Traffic_Flow_Analysis.git
cd Traffic_Flow_Analysis