Space occupancy detection
Syväjärvi, Juho (2026)
2026
Sähkötekniikan DI-ohjelma - Master's Programme in Electrical Engineering
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
Hyväksymispäivämäärä
2026-06-18
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202606077101
https://urn.fi/URN:NBN:fi:tuni-202606077101
Tiivistelmä
This thesis explored the feasibility of using Texas Instruments FMCW-radar for indoor people counting purposes. Baseline approach using the TI’s SDK provided algorithms was compared against own implementations of DBSCAN and CNN processing pipelines.
Due to the demand by the vast adoption of FMCW-radars by the automotive industry, the whole FMCW-radar system has been mainly condensed to a singular System On a Chip or SOC with antennas on the silicon chip as Antenna On Package or AOP chips. This has allowed mass production of the sensors at very low cost which has unlocked new use cases for this emerging technology. One of these is indoor and outdoor presence detection and people counting which has growing demand due to the smart building and office energy efficiency developments. For these uses knowing the amount of people in different rooms, offices, lounges or hallways is vital. Existing technologies mostly rely on PIR, TOF or camera based solutions. Each of these have their own weaknesses, for example, PIR-systems can not detect multiple persons on same room, TOF-sensors lack resolution and cameras, while providing good results have privacy, cost and processing load issues. FMCW-radars could provide good middle ground between all these with low energy consumption, high enough resolution to distinguish multiple targets and ability to operate even in total darkness all while maintaining low cost.
Fundamental theory of FMCW-radar signal processing, DBSCAN and convolutional neural networks or CNNs is explored in the early parts of the thesis with the goal of providing good and clear fundamentals of the complete radar processing pipeline while being easy to understand. The data collection was done in standard office meeting room setup with table surrounded by 5 chairs and some stools. The data was gathered by Raspberry Pi with camera. The radar provided target detections, pointcloud data and range-angle heatmap TLV messages transmitted over UART were also gathered and timestamped to a Postgres database on Raspberry Pi to be later annotated based on the automatically captured images. Large amount of time was also spent fine tuning the radar front end parameters to provide best possible results while complying with Finnish and European regulations. Ultimately large dataset consisting of over 400 hours of office space usage data was gathered, annotated and analyzed.
The results clearly indicated that, while baseline approach using the TI provided algorithms work well for presence detection, the people counting accuracy could be improved using more data driven approaches. DBSCAN provided better results than baseline after using grid search to find the optimal hyperparameter values, epsilon and minimum number of samples. The CNN approach using the computed range-time heatmaps provided better results especially at higher people counts. However the generalisation of the model can not be guaranteed without further research in addition to larger and more comprehensive dataset.
Due to the demand by the vast adoption of FMCW-radars by the automotive industry, the whole FMCW-radar system has been mainly condensed to a singular System On a Chip or SOC with antennas on the silicon chip as Antenna On Package or AOP chips. This has allowed mass production of the sensors at very low cost which has unlocked new use cases for this emerging technology. One of these is indoor and outdoor presence detection and people counting which has growing demand due to the smart building and office energy efficiency developments. For these uses knowing the amount of people in different rooms, offices, lounges or hallways is vital. Existing technologies mostly rely on PIR, TOF or camera based solutions. Each of these have their own weaknesses, for example, PIR-systems can not detect multiple persons on same room, TOF-sensors lack resolution and cameras, while providing good results have privacy, cost and processing load issues. FMCW-radars could provide good middle ground between all these with low energy consumption, high enough resolution to distinguish multiple targets and ability to operate even in total darkness all while maintaining low cost.
Fundamental theory of FMCW-radar signal processing, DBSCAN and convolutional neural networks or CNNs is explored in the early parts of the thesis with the goal of providing good and clear fundamentals of the complete radar processing pipeline while being easy to understand. The data collection was done in standard office meeting room setup with table surrounded by 5 chairs and some stools. The data was gathered by Raspberry Pi with camera. The radar provided target detections, pointcloud data and range-angle heatmap TLV messages transmitted over UART were also gathered and timestamped to a Postgres database on Raspberry Pi to be later annotated based on the automatically captured images. Large amount of time was also spent fine tuning the radar front end parameters to provide best possible results while complying with Finnish and European regulations. Ultimately large dataset consisting of over 400 hours of office space usage data was gathered, annotated and analyzed.
The results clearly indicated that, while baseline approach using the TI provided algorithms work well for presence detection, the people counting accuracy could be improved using more data driven approaches. DBSCAN provided better results than baseline after using grid search to find the optimal hyperparameter values, epsilon and minimum number of samples. The CNN approach using the computed range-time heatmaps provided better results especially at higher people counts. However the generalisation of the model can not be guaranteed without further research in addition to larger and more comprehensive dataset.