AI-Powered Traffic Light and Environmental Object Detection for Visually Impaired People

This study combines a mobile application with a machine learning pipeline to detect traffic lights and surrounding objects in real time. Camera images are processed locally and results are delivered through speech and vibration.

Method and architecture

• CNN architectures including AlexNet, VGGNet, ResNet, MobileNet and a CNN–LSTM hybrid.
• Transfer learning and fine-tuning of pretrained networks such as MobileNetV2.
• 64×64 normalization, SelectKBest, Global Average Pooling and PCA.
• On-device Android inference with TensorFlow Lite.

Application features

Real-time red/green traffic light classification, detection of nearby objects and an accessible audio interface with vibration feedback.

Reported experimental results

The study reports over 99% accuracy and a high F1 score for the CNN–LSTM model under different lighting and capture conditions, alongside positive user feedback on spoken guidance and response time.

Contributions

An end-to-end data, model and mobile pipeline; a lightweight on-device architecture; and a detection workflow combining feature selection with transfer learning.