Computer Vision-Based Dangerous Riding Behaviors Detection

Overview

E-scooters have transformed urban mobility, but reckless riding poses significant safety risks. My Master’s thesis explores the application of computer vision to detect dangerous riding behaviors, such as multiple riders on a single e-scooter.

The project was done as part of my studies at Uppsala University together with Voi Technology, a leading e-scooter sharing company in Sweden.

🔬 Research Highlights:

• Developed a YOLOv4-based object detection model, achieving 93.89% mAP for e-scooter recognition.
• Trained CNN-based image classification models, with a DenseNet-121 model reaching an F1 score of 81% in detecting unsafe rides.
• Created a custom e-scooter dataset and applied transfer learning to enhance detection accuracy.

🎯 Why It Matters:

By identifying dangerous behaviors in real time, this research contributes to improving road safety and shaping future e-scooter regulations.

📜 Read more in the full thesis