Nan Yin

In our previous study, we proposed a vision-based ranging algorithm (LRA) that utilized a monocular camera with four lasers (MC4L) for indoor positioning in dark environments. The LRA achieved a positioning error within 2.4 cm using a logarithmic regression algorithm to establish a linear relationship between the illuminated area and real distance. However, it cannot distinguish between obstacles and walls. Hence, it results in severe errors in complex environments. To address this limitation, we developed an LBP-CNNs model that combines local binary patterns (LBPs) and self-attention mechanisms. The model effectively identifies obstacles based on the laser reflectivity of different material surfaces. It reduces positioning errors to 1.27 cm and achieves an obstacle recognition accuracy of 92.3%. In this paper, we further enhance LBP-CNNs by combining it with fast Fourier transform (FFT) to create an LBP-FFT-CNNs model that significantly improves the recognition accuracy of obstacles with similar textures to 96.3% and reduces positioning errors to 0.91 cm. In addition, an inertial measurement unit (IMU) is integrated into the MC4L device (MC4L-IMU) to design an inertial-based indoor positioning algorithm. Experimental results show that the LBP-FFT-CNNs model achieves the highest determination coefficient (R² = 0.9949), outperforming LRA (R² = 0.9867) and LBP-CNN (R² = 0.9934). In addition, all models show strong stability, and the prediction standard index (PSI) values are always below 0.02. To evaluate model robustness and MC4L-IMU work reliably under different conditions, the experiments were conducted in a controlled indoor environment with different obstacle materials and lighting conditions.