Urban areas worldwide are plagued by significant traffic congestion, leading to delays, high fuel consumption, and increased emissions. Traditional traffic management systems, being static and often unresponsive to real-time conditions, further exacerbate these issues. This project aims to address these challenges by developing an innovative traffic management system that leverages deep learning to dynamically synchronise traffic signals based on real-time data. The primary objective is to optimise traffic flow, reduce travel times, and improve air quality, thereby promoting a sustainable urban environment. By evaluating multiple datasets and ultimately using GTA V for data collection, we were able to tailor our data to specific project needs and simulate precise traffic scenarios. While the implementation of signal optimization is still in progress, we have successfully collected the necessary data and established methods for future data gathering. This report details our methodology, data processing techniques, and the framework for model training and testing, laying the groundwork for a promising solution to modern traffic management challenges.
Data Collection: Camera Positioning: Virtual cameras were positioned at key intersections to capture detailed traffic data, including vehicle flow and signal timings. Recording Traffic Data: We recorded traffic footage from these strategic locations to collect data on vehicle movements and traffic signal states. Frame Extraction: Using OpenCV, we extracted frames from the recorded videos at half-second intervals, resulting in a dataset of 1450 frames. 4. Annotation and Data Processing (Rajat/Somesh): Annotation:
- Tool Used: We used Roboflow for annotating the extracted frames.
- Categories: Signals were categorised into three classes: red, green, and yellow.
- Annotation Process: Each frame was carefully annotated to identify and label traffic signals within the scene. This resulted in a total of 1556 annotations across all frames
We removed 50% of the frames containing red signal annotations to address class imbalance. This reduced the dataset to 517 frames while maintaining a representative sample of traffic signal states.
Here is a heatmap of the regions where the traffic
lights are detected.
Reasons for Choosing YOLOv8
- Speed and Accuracy: YOLOv8 maintains a robust balance between detection speed and accuracy, essential for real-time applications like traffic signal detection. It can quickly process video frames, identifying traffic signals with high precision.
- Scalability: YOLOv8 offers various model sizes, including YOLOv8n (nano), YOLOv8s (small), YOLOv8m (medium), YOLOv8l (large), and YOLOv8x (extra-large). This scalability enables us to choose a model size that best fits our computational resources and accuracy requirements.
Model Training:
- Training Configuration: We configured the YOLOv8m model with appropriate hyperparameters, including a batch size of 4 and a total of 30 epochs. This setup aimed to optimise the model’s performance while ensuring efficient use of computational resources. Baseline Performance (Initial Model)
- Precision-Confidence Curve: Green Signal: High precision with increasing confidence. Red Signal: Gradual increase in precision. Yellow Signal: Lower precision overall. All Classes: Achieves 1.00 precision at a confidence level of 0.586.
- Precision-Recall Curve: Green Signal: High precision (0.995) across recall levels. Red Signal: High precision (0.964) across recall levels. Yellow Signal: Very low precision (0.052). All Classes: Mean Average Precision (mAP) at 0.5 is 0.670.
- Recall-Confidence Curve: Green Signal: Higher recall with lower confidence. Red Signal: Gradual decrease in recall with higher confidence. Yellow Signal: Very low recall. All Classes: Recall of 0.98 at a confidence level of 0.000.
- Training and Validation Losses: Train/Box Loss: Decreases steadily. Train/Cls Loss: Decreases steadily. Train/Dfl Loss: Decreases steadily. Metrics/Precision (B): Fluctuates initially, then stabilises. Metrics/Recall (B): Fluctuates significantly, shows improvement. Metrics/mAP50 (B): Increases over epochs. Metrics/mAP50-95 (B): Steady increase.
- Confusion Matrix:
Final results