By Bernards Yvette Noah
Urban traffic congestion continues to be a major challenge across modern cities, contributing to increased travel time, fuel consumption, pollution, and accident risks. Traditional traffic monitoring methods, such as manual vehicle counting or sensor-based systems are often expensive, difficult to scale, and limited in real-time adaptability. With advances in artificial intelligence and computer vision, automated traffic monitoring using deep learning has become a practical and efficient alternative.
This project explores Traffic Flow Analysis using YOLOv8 and SORT, presenting a real-time intelligent system capable of detecting, tracking, and counting vehicles from video footage. The system integrates modern object detection and multi-object tracking techniques to generate lane-based traffic statistics that can support smarter transportation planning.
Background and Research Foundation
The project builds on existing research in deep learning–based traffic analysis. Earlier studies relied heavily on convolutional neural networks (CNNs) and two-stage detectors such as Faster R-CNN to identify vehicles in surveillance footage. These approaches provided strong detection accuracy but required high computational resources and were less suited for real-time deployment.
More recent research introduced real-time object detection frameworks such as YOLO (You Only Look Once), including implementations like YOLOv7. These models process images in a single pass, dramatically improving inference speed while maintaining competitive accuracy.
Both approaches commonly integrate tracking algorithms such as SORT (Simple Online Realtime Tracking) to maintain vehicle identities across frames. Tracking is essential because it prevents double-counting and enables extraction of motion-based metrics such as direction and speed.
The literature highlights a clear shift in intelligent transport systems: balancing detection accuracy with real-time performance for deployment using standard CCTV infrastructure rather than specialized hardware.
Problem Statement
Despite improvements in monitoring technologies, many traffic systems still lack:
- Accurate real-time vehicle detection
- Lane-specific vehicle counting
- Reliable tracking across frames
- Visual analytics for traffic flow interpretation
This project addresses these gaps by designing an automated pipeline capable of detecting vehicles, tracking them over time, assigning them to lanes, and producing real-time traffic statistics from video data.
System Design and Architecture
The proposed system combines three main modules:
1. Lane Definition Module
Lane boundaries are defined using polygon-based segmentation. Users interactively mark lane regions on a video frame, allowing the system to adapt to different road layouts.
2. Vehicle Detection and Tracking
The system detects vehicles using YOLOv8 and tracks them using SORT. Detection identifies bounding boxes and vehicle classes (cars, trucks, and buses), while tracking assigns unique IDs to vehicles across frames using Kalman filtering and IOU matching.
This combination ensures that each vehicle is counted once and its movement is monitored accurately.
3. Real-Time Visualization
Bounding boxes, confidence scores, lane assignments, and vehicle counts are displayed directly on video frames. This provides a live demonstration interface for traffic monitoring.
The implementation uses computer vision tools such as OpenCV for frame processing and visualization.
Implementation Workflow
The workflow consists of several stages:
Image Preprocessing
Video frames are resized and enhanced using techniques such as contrast adjustment and noise reduction to improve detection performance under varying lighting conditions.
Feature Extraction and Classification
The model uses pretrained weights from the widely used COCO dataset developed by Microsoft. This enables the system to recognize vehicles without requiring large custom datasets.
Vehicle Tracking
SORT maintains consistent vehicle IDs across frames using motion prediction and detection matching.
Lane Assignment and Counting
Each vehicle’s centroid is tested against lane polygons to determine its lane and update lane-specific counters.
Speed Estimation
Approximate speeds are computed using positional displacement between frames.
Key Results and Observations
The system achieved near real-time performance (approximately 20–25 frames per second on GPU hardware). Vehicles were successfully detected and tracked across multiple lanes, with clear visualization of:
- Lane-wise vehicle counts
- Vehicle classification (cars, buses, trucks)
- Confidence scores
- Approximate speeds
These results demonstrate the effectiveness of combining modern deep learning detection models with lightweight tracking algorithms for traffic analysis.
Limitations and Future Improvements
Although the system performs well, several limitations remain:
- Detection accuracy depends on camera angle and resolution
- Occlusion (vehicles blocking each other) can affect tracking
- Lane polygons must be manually defined
- Pretrained models may misclassify unusual vehicle types
Future work could include:
- Training on local traffic datasets
- Using advanced trackers such as DeepSORT or ByteTrack
- Automating lane detection
- Deploying the model on edge devices for smart city applications
Conclusion
This project demonstrates how deep learning and computer vision can transform traffic monitoring into an automated, scalable, and real-time system. By integrating YOLOv8 for detection and SORT for tracking, the system provides accurate lane-based vehicle analytics using standard video feeds.
Such intelligent traffic flow analysis solutions have strong potential to support urban planning, adaptive traffic signal control, and smart transportation systems—especially in rapidly growing cities where congestion remains a critical challenge.