Efficient and robust 3D indoor visible light positioning via uniform and power-of-two quantization on multi-head...

Existing indoor positioning methods in visible light communication systems typically rely on large databases, powerful signal processing units, and additional sensors such as gyroscopes, which limit their applicability in efficiency-sensitive scenarios. To address this issue and achieve an optimal balance between positioning accuracy, inference time, and model size, we propose uniform and power-of-two quantization on a multi-head ResNet50 (UPU-MH-ResNet50) algorithm in this paper. The proposed algorithm does not require additional sensors and can accurately estimate the three-dimensional coordinates of the receiver, including X and Y coordinates and rotation angle. It effectively mitigates the impact of random angle changes on positioning accuracy, significantly enhancing the system's robustness. Additionally, the algorithm is further optimized through model quantization to reduce computational costs. We conducted validation on a self-built experimental testbed in a laboratory environment with dimensions of 2.6m×2.6m×2.2m and a receiver rotation angle range of ±30^∘. Within this tilt range, 90% of the 3D positioning errors are controlled within 2 cm. Moreover, the proposed 4-bit UPU-MH-ResNet50 achieves a 7.7× reduction in model size and a 2.7× speedup in inference, where the acceleration and efficiency results are measured from the GPU-based implementation using real captured data collected on the developed testbed, indicating high energy efficiency.