Collaborative Multi-point User Localization with Joint Communication and Sensing

Joint Communication and Sensing (JCS) is a key feature of 6G networks. Multi-point localization using Wi-Fi Channel State Information (CSI) is a prime example of JCS. However, deploying such applications at the network edge is complex. It requires fusing CSI data collected by multiple access points (APs) and managing significant computational workloads on resource-constrained devices. This creates a two-fold challenge: 1) At the application level, how much does collaboration (using multiple APs) actually improve performance, and how robust is this performance to sensor (AP) failure? 2) At the system level, how can we deploy these collaborative applications (which have different strategies, like model splitting or data fusion) in a dynamic and efficient way, instead of using static, manual configurations? This thesis addresses both challenges. First, we implement and evaluate a complete indoor localization pipeline, we train this pipeline on a dataset of CSI traces collected from a multi-antenna, multi-anchor real-world scenario, creating two distinct versions: one for regression (predicting coordinates) and one for classification (predicting location labels). We use this pipeline to conduct a detailed robustness analysis. Our results provide a key finding: data fusion is not just beneficial, it is essential for reliable localization. We show that performance in a surveyed environment collapses from 99.99% accuracy (with 5 APs) to 55.64% (with 1 AP). Similarly, for regression task, the Mean Absolute Error (MAE) increases significantly from 3.62 cm (with 5 APs) to 73.59 cm (with 1 AP). Second, to manage such applications, we design and implement a novel distributed control-layer solution, the Collaborative Inference Manager (CIM). This is an orchestration service that uses cloud-native tools (Docker, Kubernetes) to simulate a distributed edge environment, where each container represents an independent computing node. It can take API requests and automatically deploy the CSI localization pipeline using two different strategies: one based on model partitioning and the other based on data-fusion. The contribution of this thesis is therefore two-fold: we provide both a quantitative analysis that proves the necessity of multi-point data fusion for CSI localization, and a dynamically reconfigurable framework to deploy and manage such collaborative sensing applications efficiently within a simulated environment.