An End-to-End Design Workflow for Indoor Localization on Edge Devices: From Quantized Capsule Networks to HLS Implementation

Indoor human localization has gained significant importance because of its wide range of applications, including smart environments, health monitoring and security systems. However, still one of the main challenges is having an accurate indoor positioning system due to complex environments, signal interference, and the demand for real-time processing. While traditional approaches like Recurrent Neural Network (RNNs) are often used for time-series sensor data, they can suffer from high resource consumption and difficulty in modeling complex relational patterns. To address this challenges, this thesis presents a solution using Capsule Network (CapsNet) with capacitive sensors for indoor localization and the motivation of using CapsNets lies in their inherent ability to model hierarchical part-whole relationships which allows to better interpret complex and overlapping temporal patterns from sensors compared to conventional architectures so it helps to achieve high accuracy by preserving the rich capacity of Capsule Network comes at the cost of high computational and memory requirements, making them so difficult to deploy on real-time, resource-constrained edge devices like FPGAs. This work introduces a second key contribution: a complete optimization and implementation workflow, featuring a custom mixed-precision quantization framework based on Post Training Quantization (PTQ) method and a subsequent High-Level Synthesis (HLS) implementation. The custom framework was developed to handle the unique numerical sensitivity of the CapsNet’s dynamic routing algorithm, which requires a highly accurate quantization approach. This end-to-end process successfully compressed the model by a factor 4.56x (from 59.72 KB to 13.1 KB) with negligible accuracy degradation (0.5% on test set and less than 2% tatally), demonstrating the feasibility of deploying this advanced architecture in practical applications. This thesis provides a complete methodology for designing and implementing a novel deep learning model for efficient and accurate indoor localization on edge devices. Despite these successful results of our work, several key challenges due to complexity and high sensitivity of CapsNet remain. The high numerical sensitivity of the dynamic routing algorithm, which requires maintaining high-precision (18-20 bit) activations to preserve accuracy, poses a significant hurdle for further optimization. Future work should focus on Quantization-Aware Training (QAT), which could make the model inherently more robust to aggressive, low-bit quantization because it allows the model to be trained on simulated quantized data so the model learn to adept and handle the constraints of low-precision data. Furthermore, a complete hardware resource and power analysis on a target FPGA is required to validate real-world efficiency. Exploring architectural enhancements to the Capsule Network itself, such as alternative routing algorithms, also presents a promising avenue for improving both performance and efficiency. This thesis provides a complete methodology for designing, optimizing, and implementing a novel deep learning model for efficient and accurate indoor localization on edge devices.