Presence Sensor Projects

Summary

This webpage describes my work at a Retailtech startup company. In Retailtech, we aimed to find one of the most obvious, but at the same time previously unobtainable metrics — foot traffic at a retail spot. The retail store is a challenging place to track position, because the most accurate method so far, GNSS, obviously doesn't work indoors. Some startups tried to track shoppers in the store using Ultra-Wideband technology, which provides accurate positioning, though installations always turned out to be expensive, because a large number of location beacons had to be installed. Some tried to use Bluetooth and WiFi positioning, but they also require many beacons and have lower accuracy (2–5 m), which is unacceptable.

At Retailtech, a different approach was chosen: for each spot of interest, a sensor was installed that tracks traffic locally. The sensor is a distance sensor, which constantly measures the distance, waits for a change in readings, and considers that a customer arrived. Based on how close a customer is and how long the stay lasted, we decided whether the person took a product, was interested but didn't buy, or was just passing by.

Even though this approach gave good results in most cases, there were some scenarios where it had false readings: for example, a shop employee restocking shelves or a person looking at products at a shelf in front of the spot of interest. Therefore, the decision to replace the sensor with a camera was made.

Sensor Versions

Original Sensor

This is the original sensor from when I first arrived at Retailtech.

A cost-effective solution based on the HC-SR04 ultrasonic distance sensor and ESP32 for WiFi connectivity.
It was equipped with a buzzer, green and red LEDs, and two buttons.
Unfortunately, it was too bulky, had a cumbersome setup procedure, and had lower performance with distances over 2 m to the target.
I was in charge of modifying the existing firmware and creating a dashboard on the web.

Prototype 2 — Infrared Time-of-Flight (ToF) Sensor

An upgraded version of the previous sensor that I developed.

I was responsible for developing the electronics and firmware.
Enhanced performance and accuracy by around 2–5× with a smaller size for easier installation.
Equipped with a VL53L1X ToF sensor, a smaller buzzer, an RGB LED, and one button.
Designed the enclosure with functionality and manufacturing in mind for easy assembly.
Allowed the sensor to be placed near the price tag, which is a great location for collecting readings.

Prototype 3 — AI Camera Based on Raspberry Pi

A camera sensor made by our partners, connected using WiFi.

I was in charge of updating the firmware.
Raspberry Pi Compute Module 4–based design.
Used proprietary AI software to detect person presence and movement.
The design allowed easy placement of the sensor at the retail spot.
Later, I rewrote it to use an open-source YOLO model to detect people, hands, and faces.

Prototype 4 — Final and Unfinished Prototype Based on Raspberry Pi with CAT-M 4G Connectivity

I was responsible for developing the electronic design and firmware.
Updated design made for installation in coolers for customers.
Based on customer requests, sensors were equipped with a 4G module and SIM cards for better connectivity.
Had camera port, HDMI, USB, CAT-M, GPIO, and optional GNSS for outdoor use.