Congestion Management Study Pairing Drone Images with Simulated Traffic Models Using Real Data in Alabama Offers Lessons for Congestion Mitigation.
Traffic monitoring is an important part of congestion mitigation and traffic management. This study proposed congestion mitigation approach based on the premise that a short interval analysis (one to five minutes) would be sufficient to manage congestion defined by exceeding the road capacity and is characterized by high vehicular density and vehicle speeds that are lower than the desired speeds. A number of interconnected, self-managed drones were utilized to establish a deployable system to perform immediate monitoring and assessment for traffic conditions to infer if congestion is approached. In order to detect vehicles, a machine learning technique of Convolutional Neural Network (CNN) was utilized that used a single neural network to the full image which divided the image into regions and predicted bounding boxes and probabilities for each region. Two neural networks were trained using only the five desired classes to be detected: car, bus, truck, motorbike, and bicycle. Drone cameras were calibrated using real values observed in Birmingham, AL in 2019, compared to their apparent values in images to track any detected vehicles. The features correlated to traffic congestion were reproduced utilizing a traffic simulation model and the proposed methodology was tested by collecting and investigating video images from drones. In addition, detection limitations of drones were tested using multiple videos from Raleigh, NC as well as Georgia DOT to estimate the minimum size of detection.
- Test drones’ vehicle detection limitations prior to deployment. Results from this study pointed out that the bounding box of the vehicle must have the size of 100 pixels at least, and its smallest dimension must not be less than 8 pixels for accurate vehicle detection.
- Deploy drones prior to actual event requiring traffic management for better validation of vehicle detection and calibration of the traffic stream models. For locations that are pre-selected for drone array deployment, it would be advisable to deploy the drones prior to the actual event requiring traffic management to validate the detection and calibrate the traffic stream models if they will be used for traffic state assessment.
- Use a Graphics Processing Unit (GPU) to accelerate vehicle detection. In this study, increasing the number of processing cores led to less detection time. Therefore, GPU was used to accelerate detection since GPU has a higher number of smaller cores and it can carry out parallel processes more efficiently.