TLDR: DARTS is a drone-based, AI-powered system that uses thermal cameras and deep learning to detect traffic incidents and monitor congestion in real-time. It achieved 99% accuracy and detected a real-world crash 12 minutes faster than traditional methods, offering a flexible, scalable, and cost-effective solution for improving road safety and traffic management.
Road traffic incidents are a major global concern, leading to millions of fatalities and injuries annually, alongside significant traffic congestion and secondary crashes. Traditional methods for detecting these incidents, such as closed-circuit television (CCTV), dashcam footage, and sensor-based systems, often suffer from limitations. These include restricted flexibility, a need for extensive infrastructure, poor performance in low-visibility conditions like nighttime or fog, and potential privacy issues due to visible-light cameras. These drawbacks hinder their adaptability and scalability, especially in rapidly changing incident hotspots.
Introducing DARTS: A New Era in Traffic Surveillance
To address these critical challenges, researchers have developed DARTS: a Drone-Based AI-Powered Real-Time Traffic Incident Detection System. This innovative system integrates several advanced technologies to provide a more efficient and reliable solution for traffic management. DARTS leverages the high mobility and aerial perspective of drones for adaptive surveillance, incorporates thermal imaging for superior performance in low-visibility conditions and enhanced privacy protection, and utilizes a lightweight deep learning framework for real-time vehicle trajectory extraction and incident detection.
The system has demonstrated impressive capabilities, achieving a 99% detection accuracy on a specially collected dataset. Beyond just detection, DARTS supports simultaneous online visual verification, allows for severity assessment of incidents, and monitors the propagation of incident-induced congestion through a user-friendly web-based interface. For a deeper dive into the technical details, you can read the full research paper here.
How DARTS Works
The development of DARTS involved a meticulous methodology, starting with data collection. High-quality thermal traffic monitoring videos were gathered from freeways under various conditions—incidents, recurrent congestion, and normal traffic—using drones flying at optimal altitudes and speeds. These continuous video streams are then segmented into 2-minute clips for processing.
The core of DARTS is its traffic incident detection framework, which consists of four interdependent components:
- Trajectory Image Generation: This component identifies vehicles in thermal video segments, extracts their movement trajectories, and generates trajectory images at fixed intervals. This process uses a custom-trained YOLO model for vehicle detection and the Lucas-Kanade optical flow tracker for movement tracking, with motion compensation to account for the drone’s movement.
- Traffic Incident Detection Model (TCD-Net): Built upon the generated trajectory images, TCD-Net is a deep learning model based on a Convolutional Neural Network (CNN) architecture. It’s enhanced with multiscale CNN structures, a Convolutional Block Attention Module (CBAM), and Spatial Pyramid Pooling (SPP) to effectively extract traffic features. This allows it to classify traffic conditions into incident, recurrent congestion, or normal traffic, accurately distinguishing between them. The model performed best with monochrome images and a 20-second trajectory extraction period.
- Image-to-Video Aggregation: After classifying individual trajectory images, a statistical aggregation method is used to convert these image-level detections into a single, comprehensive video-level incident detection result for each 2-minute segment. This method helps to mitigate the impact of any isolated misclassifications.
- Incident Feature Extraction: Once an incident is detected, DARTS extracts crucial features such as the incident scene’s time period, the length of the incident-induced non-recurrent congestion, and its propagation speed. This information is vital for transportation management centers (TMCs) to assess the incident’s impact.
The entire system is integrated into a traffic incident detection platform, which includes both drone-side software (an Android application for data forwarding) and workstation-side software (a web-based application with an interactive Graphical User Interface). This platform supports real-time incident detection, automated feature extraction, and web-based visualization of results, allowing TMC personnel to monitor and manage incidents effectively.
Also Read:
- AI Agents Learn to Optimize Urban Traffic Flow
- Spatiotemporal Error Adjustment Enhances Deep Learning Traffic Models
Real-World Impact and Future Outlook
A field test conducted on Interstate 75 in Florida provided compelling validation of DARTS’s efficacy. During the test, the system successfully detected a rear-end collision and accurately tracked the resulting non-recurrent congestion. Crucially, DARTS detected the crash at approximately 5:03 PM, which was 12 minutes earlier than the local TMC’s reported time of 5:05 PM and verification time of 5:15 PM. This early detection, combined with immediate online access to the incident scene for verification, has significant implications for accelerating emergency response times and potentially reducing severe injuries and fatalities.
The study also highlights several forward-looking contributions. It introduces the first publicly available drone-based thermal traffic monitoring dataset, which is a critical resource for future research. Furthermore, the computational efficiency and modular design of DARTS suggest its potential for deployment in distributed, multi-drone patrolling systems, possibly running on edge computing devices directly on the drones. This vision is supported by evolving regulatory frameworks for Beyond-Visual-Line-Of-Sight (BVLOS) operations and advancements in drone docking stations for automated operations.
DARTS offers enhanced operational flexibility compared to traditional fixed surveillance systems, allowing for dynamic adaptation of patrolling schedules and routes. Its ability to cover a relatively large area per flight cycle indicates scalability for wider freeway networks or remote regions. By reducing the computational load at TMCs and offering a cost-effective, infrastructure-light solution, DARTS holds particular promise for regions with limited conventional infrastructure, contributing to more inclusive and scalable incident management strategies and advancing global road safety and sustainable transport objectives.
