A New Blind Road Classification Method Based on Wavelet Feature Fusion and Transfer Learning for Mobile Environment

It's very useful to sense blind road environment, a new real-time detection for blind road detection by wavelet feature fusion and transfer learning based on MobileViT is proposed. Firstly, the wavelet-based features are extracted through wavelet transformation with different scales and channels, some bottleneck features are recovered from pretrained model in other big-data domain, and then these features are incorporated together. Secondly, the new model is trained again based on pretrained model parameters with fine tuning. Finally, we used the trained model for real time blind road scene detection. In our experiments, the blind road image datasets are collected at different districts in Chengdu for tests with different model parameters. At the same time, our inference models are tested with the embedded device in real time environment. Our test results show that the new method has advantages on classification precision with directly using MobileViT by transfer learning. The proposed scheme is more effective than the original scheme, and the total accuracy is increased from 88.09% to 91.58% on the standard flower photos dataset. We got the total precision of 99% and 95% respectively on two blind road datasets collected by ourself.

1. Introduction

It is not easy for the blind people to move around freely and safely, which makes it difficult for them to attend in public and professional exercises outdoors and indoors. As reported by WHO (World Health Organization)'s statistical analysis, more than 285 million people are blind or with amblyopia in our world. Among those, 246 million people suffer from extremely difficult visual impairment [1].