• Citer

Passive radar based on Wi-Fi signals for crowd monitoring: From multi-target tracking to large-scale crowd analysis

par Storrer, Laurent
Président du jury De Doncker, Philippe
Promoteur Horlin, François ;Pollin, Sofie
Publication Non publié, 2023-12-05

Thèse de doctorat

• Accès en ligne
• Détails
• Contenu
• Statistiques

Résumé :

We explore the potential of passive radars based on Wi-Fi signals for crowd monitoring in mass gatherings and pedestrian mobility. Crowd monitoring aims at avoiding excessive people density to prevent accidents, and at analysing people's mobility patterns. It requires privacy-friendly people counting and trajectory estimation techniques, for which radars are especially well-suited. In particular, passive radars based on the ubiquitous Wi-Fi signals reuse channel estimation fields for sensing purposes. We address the problem of static clutter removal and detect slowly moving targets in highly cluttered environments. We show that simple novel methods such as Average Removal are efficient alternatives to the computationally intensive Extended Cancellation Algorithm. We propose a multi-antenna experimental setup based on software-defined radios, and relying on channel estimation signals from the latest Wi-Fi standard, namely 802.11ax, with 80 MHz of bandwidth. We design and experimentally validate a solution for the calibration of hardware-induced phase shifts between signals at different antennas using an anchor. We apply Angle-of-Arrival estimation for multiple people on detected cells in measured range-Doppler maps. This enables the implementation of a multi-people tracking scheme based on a Joint Probabilistic Data Association Filter combined with an Unscented Kalman Filter. This scheme is tuned according to the heavy constraints imposed by the Wi-Fi-based passive radar, and validated experimentally indoors, proving the feasibility of full range-Doppler-angle passive radars for people trajectory tracking with 802.11ax signals. We extend the analysis towards large-scale crowds, first by proposing a novel full-wave crowd simulator, through the resolution of an electric field integral equation applied to human bodies in two dimensions using the Method of Moments (MoM). This enables simulation with people in the near-field region of each other. We demonstrate that it can be computationally optimised by exploiting the block low-rank system structure, enabling the simulation of range-Doppler maps with up to 100 people. Then, we develop a crowd counting framework as a classification problem. A first approach resorts to Support Vector Machines applied on features extracted from range and Doppler 1D profiles. A second approach exploits Convolutional Neural Networks operating on range-Doppler maps. We perform crowd counting with high accuracy on real-life experimental measurements of a crowd with up to 100 people, where the hyperparameters of the classifiers have been tuned with MoM-based simulation data. This again proves the viability of 802.11ax-based passive radars for crowd monitoring. We also show that classifiers trained only on MoM-based simulation data can be applied to real-life measurement data, although for a reduced number of people. Finally, we explore a new track. We analyse people's flow in a street by estimating the number of people through the partition of the range-Doppler map into the negative and positive speeds, to which our counting scheme is combined with an estimation of the average people's speed.