Deep Learning Based Hybrid Vision Method for Tunnel Liner Falling Block Recognition Research
Haoxuan Qi · 2024
This paper presents an innovative approach to enhancing video surveillance and sound source localization within tunnel environments, specifically focusing on the detection of falling blocks. We introduce a robust background subtraction algorithm, the Adaptive Mean and Gaussian Background Subtraction (AMGBS) algorithm, which utilizes grayscale stability intervals and average grayscale values to optimize traditional methods. Additionally, we explore the application of deep neural networks for background modeling, leveraging their ability to learn from training images automatically. A multi-mode background subtraction method is also proposed, which adapts to various lighting conditions through the use of RGB and YCbCr color spaces. To complement visual data, we propose a Time Difference of Arrival (TDOA)-based algorithm for sound source localization, which is particularly effective in tunnel environments where visual detection may be impaired by lighting or obstructions. Through extensive experimentation, including simulated tunnel blockage scenarios, our methods demonstrate significant improvements in accuracy and reliability over existing techniques.