# A semi-trailer truck right-hook turn blind spot alert system for detecting vulnerable road users using transfer learning

**Authors:** Charles Tang  
**arXiv:** [2303.11223](https://arxiv.org/abs/2303.11223)  
**Date:** 2023-01-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.11223v1/

## Key points

- Cycling is becoming increasingly popular as a sustainable and healthy mode of transportation.
- Cyclists face significant risks when sharing the road with semi-trailer trucks.
- Right-hook turns and drivers' lack of attention to blind spots contribute to truck-cyclist collisions.
- A visual-based blind spot warning system was developed using deep learning techniques to detect cyclists for semi-trailer truck drivers.
- Several cyclist detection models were created, achieving over 90% mean average precision (mAP).
- The object detection model was deployed on a Google Coral Dev Board mini-computer with a camera module, achieving inference times as low as 15 milliseconds.
- The end-to-end blind spot cyclist detection device accurately and rapidly detects cyclists in real-time traffic scenarios.
- This device has the potential to significantly improve cyclist safety.
- Future studies could focus on implementing this device in the trucking industry and assessing its long-term impact on cyclist safety.

## Summary

Cycling is becoming increasingly popular as a sustainable and healthy mode of transportation. However, cyclists face significant risks when sharing the road with semi-trailer trucks due to truck-cyclist collisions caused by right-hook turns and drivers' lack of attention to blind spots. To address this issue, a visual-based blind spot warning system was developed using deep learning techniques to detect cyclists for semi-trailer truck drivers. The study involved creating several cyclist detection models with over 90% mean average precision (mAP), including the EfficientDet Lite 1 and SSD MobileNetV2 architectures. These models were fine-tuned on a newly proposed cyclist image dataset comprising more than 20,000 diverse images. The object detection model was then deployed on a Google Coral Dev Board mini-computer with a camera module to analyze its speed performance, achieving inference times as low as 15 milliseconds. In real-time traffic scenarios, the end-to-end blind spot cyclist detection device was tested and further analyzed for its performance and feasibility. The results indicated that this portable blind spot alert device can accurately and rapidly detect cyclists, potentially leading to significant improvements in cyclist safety. Future studies could focus on determining the feasibility of implementing this proposed device in the trucking industry and assessing its long-term impact on cyclist safety. Overall, this research presents an innovative solution to reduce truck-cyclist collisions by providing timely warnings to semi-trailer truck drivers about cyclists in their blind spots.