Solar cells are usually used as power source, but can be used for sensing as well. We propose a novel indoor tracking system that tracks people by using the change in output caused by their shadows. First, we develop a simulator to determine how the solar cells should be positioned in the tracking environment to get the best detection rate. This applies ray tracing in a model of the environment, and uses the standard deviation of solar cell output to compare different positions. Next, we apply changepoint detection methods to convert the solar cell output to a binary signal. One approach uses Bayesian online changepoint detection and another uses the change of gradient in the signal. Finally, the binary output from multiple solar cells is fused to track multiple targets in a real indoor environment in different scenarios. For this, we have implemented a particle filter based on the probability hypothesis density filter. We compare this with a tracking algorithm that uses a hidden Markov model. We have combined everything to show that it is possible to track up to two people in an indoor environment in different lighting conditions using a network of multiple solar cells.

A Camera Submodule for the Environment Observation Module for a Zebro robot was designed. It is able to collect and process data about obstacles on a Raspberry Pi and send this to the main processor of the Environment Observation module using UART.
The two main functions that were implemented are a laser rangefinder based on triangulation and a motion detector based on optical flow. The laser rangefinder calculates the distance to an obstacle by finding its own laser point in a frame. Testing showed that it is especially accurate at short distances (0-4 cm), and the range can differ between 0.30 to 6m due to environmental light and obstacle color.
The motion detector uses dense optical flow data that is provided by the h264 encoder of the Pi Camera. This data is filtered and moving obstacles are grouped together. In the tests, a moving obstacle was detected every time, but 29% also showed false positives due to large egomotion.