Blind-Spot Collision Detection System for Commercial Vehicles Using Multi Deep CNN Architecture.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Road Safety Engineering

Background:

• Heavy vehicles possess larger blind spots than cars, increasing the risk of severe accidents.
• Current vision-based object detection models often rely on single feature descriptors, limiting their effectiveness.

Purpose of the Study:

• To propose novel convolutional neural network (CNN) designs for detecting blind-spot collisions in heavy vehicles.
• To enhance object detection by integrating high-level feature descriptors and a fusion approach.

Main Methods:

• Developed two CNNs utilizing high-level feature descriptors and integrated them with the faster R-CNN framework.
• Implemented a fusion approach combining pre-trained Resnet 50 and Resnet 101 networks for feature extraction.
• Validated the models on a custom bus blind-spot dataset and the public LISA dataset.

Main Results:

• The proposed fusion approach significantly improved the performance of faster R-CNN for blind-spot detection.
• Achieved low false detection rates (3.05% and 3.49%) on the custom dataset, demonstrating suitability for real-time applications.
• Outperformed existing state-of-the-art methods in blind-spot vehicle detection.

Conclusions:

• The integrated CNN and feature fusion approach offers a robust solution for real-time blind-spot collision detection in heavy vehicles.
• This advancement has the potential to significantly reduce accidents and improve safety for pedestrians and cyclists around large vehicles.