Privacy-preserving Monitoring of Customer Flows and Intentions

AI Quick Summary

The proposed study addresses the SBB challenge of optimizing passenger flow by developing a privacy-preserving edge computing solution that analyzes and anonymizes real-time passenger data. The solution will ensure data privacy by aggregating, masking, and adding noise to the data, collecting only essential movement information while discarding identifying details. Focused use cases include pose detection and tracking for holistic passenger flow modeling and human/social intent prediction for detecting dangerous situations on train platforms. Building on our expertise in real-time edge computing with machine learning models on Xilinx FPGA SoC devices, we plan to collaborate with the Institute of Sensors and Electronics at the University of Applied Sciences Northwestern Switzerland, leveraging their advanced machine learning capabilities to refine and implement the solution. The proposed idea addresses the SBB challenge of "Optimizing Passenger Flow." Vision and JustificationSBB would like to understand better passenger flows along their passengers' multimodal journeys to address a variety of objectives and issues:Enhance Safety on the boarding platforms, e.g., by reducing "crowdedness."Improve Customer Journey/Experience, e.g., by providing better routing between connecting trains or across train stations.Increase "Situational Awareness" to adapt to live events, e.g., by predicting future customer flows.Boost Business visibility and Foot Traffic for businesses renting space at the train stations, e.g., by increasing business visibility and traffic.Validate Improvement Campaigns by increasing the understanding of the actual passenger flows to verify if campaigns or improvements have the desired effect (baselining).Today, SBB already has some observation capabilities (CCTV, Bluetooth beacons, surveys, etc.), but comprehensive customer flow monitoring is complex and often triggers privacy concerns. For example, a data analysis campaign to gain better passenger flow insights launched a few years ago (https://news.sbb.ch/artikel/116081/personenfluesse-in-bahnhoefen-besser-kennen-worum-es-geht) was met with public scrutiny due to privacy concerns and had to be adapted. Moreover, every approach has limitations (data quality and completeness, quality of privacy features, real-time applicability, etc.). Thus, comprehensive modeling capabilities of multimodal passenger flows are likely only possible with a combined set of privacy-protecting data collection and analysis methods.Proposed Idea: Innovation in Privacy-Preserving Data CollectionWe propose building and evaluating a privacy-preserving edge computing solution that analyzes passenger information in real-time, e.g., from video streams, then anonymizes the data (aggregation, masking, adding noise, etc.), and finally collects those data for further analysis. The solution is privacy-preserving ("privacy by design") since no data "leaving" the edge device can be used to identify one single person. For example, this could mean that only information on passenger movements (from, to, speed, timestamp, etc.) is collected, and all other data are discarded.Likely, the scope will need to be refined and focus on one particular (and more narrow) use case as a proof-of-concept, like:Using pose detection and tracking to collect information to build a holistic passenger flow model.A combination of pose detection and tracking with human/social intent prediction is used to observe movements on train platforms to detect dangerous situations that could be used to warn passengers and/or train operations.The primary project objective would be evaluating what data can be collected that will benefit the selected use case while preserving the privacy of the observed passengers.Our Background: Expertise in Real-Time Edge ComputingIn previous projects, we have run machine learning models on edge devices, e.g., Xilinx FPGA SoC (System-on-Chip), to energy-efficiently and autonomously detect and track objects in a real-time video feed (e.g., https://www.ateleris.ch/portfolio/machine-learning-space/). While energy efficiency may not be an issue for this idea, these solutions allow for the detection to be executed at the data-taking location, e.g., next to the camera, and can, therefore, obfuscate and anonymize the data stream at the source. An example of a video anonymization service can be seen here: https://bazar-object-detection.ateleris.com/.CollaborationWe propose collaborating with a partner, the Institute of Sensors and Electronics (ISE) at the University of Applied Sciences Northwestern Switzerland (FHNW). We have already collaborated in the past on research studies like the one previously mentioned on detecting objects in real-time on accelerated hardware. The ISE team has a Master's student working on implementing more sophisticated machine learning models and training processes on embedded hardware (primarily Xilinx). She would be able to work on the project. Our contribution could be to hardware and sensor integration, support for anonymization processes, and data analysis.

Project Idea Metadata

• Project Idea Name: Privacy-preserving Monitoring of Customer Flows and Intentions
• Date: 3/16/2025 7:30:10 PM
• Administrators:
  • Laszlo Etesi

Project Idea Description

The proposed idea addresses the SBB challenge of "Optimizing Passenger Flow."

Vision and Justification

SBB would like to understand better passenger flows along their passengers' multimodal journeys to address a variety of objectives and issues:

• Enhance Safety on the boarding platforms, e.g., by reducing "crowdedness."
• Improve Customer Journey/Experience, e.g., by providing better routing between connecting trains or across train stations.
• Increase "Situational Awareness" to adapt to live events, e.g., by predicting future customer flows.
• Boost Business visibility and Foot Traffic for businesses renting space at the train stations, e.g., by increasing business visibility and traffic.
• Validate Improvement Campaigns by increasing the understanding of the actual passenger flows to verify if campaigns or improvements have the desired effect (baselining).

Today, SBB already has some observation capabilities (CCTV, Bluetooth beacons, surveys, etc.), but comprehensive customer flow monitoring is complex and often triggers privacy concerns. For example, a data analysis campaign to gain better passenger flow insights launched a few years ago (https://news.sbb.ch/artikel/116081/personenfluesse-in-bahnhoefen-besser-kennen-worum-es-geht) was met with public scrutiny due to privacy concerns and had to be adapted. Moreover, every approach has limitations (data quality and completeness, quality of privacy features, real-time applicability, etc.). Thus, comprehensive modeling capabilities of multimodal passenger flows are likely only possible with a combined set of privacy-protecting data collection and analysis methods.

Proposed Idea: Innovation in Privacy-Preserving Data Collection

We propose building and evaluating a privacy-preserving edge computing solution that analyzes passenger information in real-time, e.g., from video streams, then anonymizes the data (aggregation, masking, adding noise, etc.), and finally collects those data for further analysis. The solution is privacy-preserving ("privacy by design") since no data "leaving" the edge device can be used to identify one single person. For example, this could mean that only information on passenger movements (from, to, speed, timestamp, etc.) is collected, and all other data are discarded.

Likely, the scope will need to be refined and focus on one particular (and more narrow) use case as a proof-of-concept, like:

• Using pose detection and tracking to collect information to build a holistic passenger flow model.
• A combination of pose detection and tracking with human/social intent prediction is used to observe movements on train platforms to detect dangerous situations that could be used to warn passengers and/or train operations.

The primary project objective would be evaluating what data can be collected that will benefit the selected use case while preserving the privacy of the observed passengers.

Our Background: Expertise in Real-Time Edge Computing

In previous projects, we have run machine learning models on edge devices, e.g., Xilinx FPGA SoC (System-on-Chip), to energy-efficiently and autonomously detect and track objects in a real-time video feed (e.g., https://www.ateleris.ch/portfolio/machine-learning-space/). While energy efficiency may not be an issue for this idea, these solutions allow for the detection to be executed at the data-taking location, e.g., next to the camera, and can, therefore, obfuscate and anonymize the data stream at the source. An example of a video anonymization service can be seen here: https://bazar-object-detection.ateleris.com/.

Collaboration

We propose collaborating with a partner, the Institute of Sensors and Electronics (ISE) at the University of Applied Sciences Northwestern Switzerland (FHNW). We have already collaborated in the past on research studies like the one previously mentioned on detecting objects in real-time on accelerated hardware. The ISE team has a Master's student working on implementing more sophisticated machine learning models and training processes on embedded hardware (primarily Xilinx). She would be able to work on the project. Our contribution could be to hardware and sensor integration, support for anonymization processes, and data analysis.

The proposed study addresses the SBB challenge of optimizing passenger flow by developing a privacy-preserving edge computing solution that analyzes and anonymizes real-time passenger data. The solution will ensure data privacy by aggregating, masking, and adding noise to the data, collecting only essential movement information while discarding identifying details. Focused use cases include pose detection and tracking for holistic passenger flow modeling and human/social intent prediction for detecting dangerous situations on train platforms. Building on our expertise in real-time edge computing with machine learning models on Xilinx FPGA SoC devices, we plan to collaborate with the Institute of Sensors and Electronics at the University of Applied Sciences Northwestern Switzerland, leveraging their advanced machine learning capabilities to refine and implement the solution.