I am Nicolas Baumann

Every cat owner knows the problem of his or her cat returning into the house from outside with prey. This leads to a time consuming cleaning effort. A solution to this problem is, to employ state-of-the-art Computer Vision (CV) techniques and utilising Convolutional Neural Networks (CNN) to detect, if a cat wants to enter the catflap with prey. However, cats are only expected to enter the catflap with prey 3% of times, which leads to a largely imbalanced classification problem. A custom and scalable image data gathering network has been built, to simplify and maximize the collection of training data. It features multiple distributed Camera Nodes(CN), a centralized master archive and a custom labeling tool. As a result of the data gathering network, 40 GB of training data have been amassed.This thesis exploits Transfer Learning (TL) methods to generalize the problem, such that it is scalable and applicable to any cat in any environment. Further, the implemented system manages to run on an off-the-shelf Raspberry Pi 4 (RPI4) at an average detection frame rate of 1 FPS. This is achieved by an asynchronous queue that dynamically adjusts the processing rate of the queue, if a cat appears on an image. The mean time, that a cat is expected to wait for the cascade to evaluate, if the cat has a prey in its snout, is 9.6 s.We show, that by arranging a cascade of TL models a recall rate of 93.3% with a FalsePositive Rate (FPR) of 28.5% can be achieved. Meaning that the cascade will correctly identify 93.3% of the cats prey entries, while falsely locking out the cat 28.5% of the times that it enters without prey.Consequently this thesis covers; a CNN cascade with its data gathering and training process, that is scalable and able to generally classify any cat, if it has a prey in its snout; at a high recall rate with low FPR; while being edge implemented on an off-the-shelf processing unit with minimal time-overhead. Featured on: Offcial Raspberry Pi Blog 10th July 2020

The main goal for this thesis is to design and to implement a ultra-low-power, self-sustaining, high precision, wrist-worn variometer with a minimal formfactor but infinite lifetime, which can display and indicate the vertical velocity. Thus resulting in a self sustaining smart watch with paragliding capabilities. Experimental results show a power consumption of 17.12 uW in harvesting mode and 1940.81 uW while running full system load (worst case scenario). The system is aided by a small rechargeable battery which acts as a buffer to either save or to power the system whenever needed. The system can power itself using the battery alone under worst case scenario (implies a paragliding flight) for a duration of 380 hours. Taking into account that such a flight might last for a maximum of 12 hours and that the system then retains into it's harvesting mode to minimize power drain and simultaneously recharge the energy-buffer, it is safe to say that the Piepser V2 is a self sustaining system and can be used in a real life scenario for as long as the lifetime of it's electrical components without having to recharge. The self sustaining aspect is being handled by the thermal dissipation of the human wrist and solar energy harvesting of the wrist worn device. The Piepser V2 is physically built into the case of the Matrixindustries Powerwatch, such that we can utilize their thermal harvesting cooler and not waste too much time on the physical design. The wearable variometer is designed around the novel Ambiq Apollo2 micro-controller, a commercial high accuracy MEMS sensor and a ultra-low-power memory-in-pixel display. Published in IEEE Transactions on Instrumentation and Measurement ( Volume: 69 , Issue: 4 , April 2020).

The goal of this project is to develop an ultra-low-power standalone navigationdevice that helps you navigate in narrow places (e.g. Niederd ̈orfli in Z ̈urich)where services like ”Google Maps” may not propose the optimal path becausethey don’t know every single small alley. Our device consists of a GPS, amagentometer and a small display that shows an arrow pointing in the directionyou should be heading, like the compass of ”Jack Sparrow” in ”The Pirates ofthe Carribbean” that points to the place you most desire.

The goal or the motivation for paraglider pilots is to stay as long in the air as possible. Therefore paraglid- er pilots are always searching for thermal upwinds that al- low them to gain altitude. These thermal lifts are difficult to detect as we cannot see them with our eyes and we can only feel the vertical acceleration but not the vertical speed with our body. Therefore devices which indicate the vertical speed (so called Variometers) are widely used among paraglider pilots. The main goal for this paper is to design and to implement a low-power, low-cost and wearable variometer with a minimal formfactor but long lifetime, which can display and indicate the vertical velocity. Experimental results show a power consumption of 6.27mW at full load. We managed to achieve a price point of 30.2 CHF per unit and a device volume of 47x47x25 [mm] at a weight of 43 grams. The wearable variometer is designed around the novel Ambiq Apollo2 microcontroller, a commercial high accuracy MEMS sensor and includes a software Kalman filter. Published in: 2019 IEEE Sensors Applications Symposium (SAS)