Introduction
The intricate tapestry of urban life is often woven with the threads of a vexing predicament: traffic congestion. This pervasive issue exacts a toll in the form of diminished productivity, heightened pollution levels, and the collective exasperation of commuters. The bustling global capital Bangkok is no exception, with traffic congestion contributing to productivity losses, increased air pollution levels, and widespread frustration among residents.
Acknowledging that significant reductions in commuting time may primarily require alterations to road layouts and routes, smaller-scale enhancements can be realized through optimizing the current network. AI-driven Traffic Management for a Better Bangkok (ATMBB) is designed to serve as a solution to this challenge.
Proposed Solution
The proof-of-concept developed features a traffic signal calibration system that utilizes footage from the existing network of surveillance cameras. This footage is mapped, labeled, and processed through a Machine Learning model to extract detailed information, such as the number of vehicles queuing at each side of an intersection. A visualization dashboard has been implemented, enabling relevant authorities to monitor current signal timings and detection data. The dashboard will also include heatmaps to provide a clearer understanding of traffic density in specific areas.
Methodology
Dataset Acquisition
We utilized and leveraged several datasets, which were derived from the CCTV system as part of the Samyan initiative, led by Chulalongkorn University's Property Management sector (PMCU).
Pre-processing
The datasets were initially gathered and subsequently labeled using the Roboflow platform. Python was employed to process the raw footage, extracting frames, and storing the results. Upon completion of the annotation, Roboflow provided the option to specify the target model type and export the dataset in the corresponding format. The final step involved merging the segmented datasets to create the foundational dataset.
Results
We initially reproduced a YOLO model, achieving a mean confidence score of approximately 70% using the base version 5 model without any additional algorithms. To enhance both the detection rate and confidence score, we first applied the SAHI algorithm. Subsequently, we implemented DeepSORT to address the issue of duplication caused by occlusion.
Summary of Accomplishments
We successfully developed a vehicle detection model with a +70% accuracy. Additionally, we enhanced the system to derive both car counts and velocity, and we constructed a heatmap. Essentially, we created a proof-of-concept AI-driven traffic management platform that incorporates a traffic signal calibration system that utilizes surveillance footage, mapping, and labeled data to train a Machine Learning model for precise vehicle counting at intersections. The data is then used to calculate optimal signal timings. The system's dashboard provides real-time displays of traffic signal timings, vehicle counts, and traffic density heatmaps, offering valuable insights for traffic management.
Future Directions
Several key functionalities, integral to the minimal viable product, remain under development by our team: the signal control method, signal optimization model, and advanced 2D simulated visualizations. Additionally, as we expand the base model, latency is likely to increase, potentially necessitating a migration to one of NVIDIA's existing models or frameworks.
