Gunshot detection plays a critical role in protecting African wildlife and reducing illegal poaching activities. The decline of keystone species disrupts ecosystems and inhibits forest CO₂ absorption, contributing to climate change. To support conservation efforts, this thesis contributes to the development of an embedded acoustic surveillance system that detects gunshot sounds and locates their sources, enabling rangers to respond in real time. However, realizing such a system is challenging due to limitations in data, resources, and infrastructure.

This thesis proposes a lightweight convolutional recurrent neural network, combining depthwise separable convolutions (DSConv) and a gated recurrent unit (GRU), designed to detect the onset of gunshot sounds for trilaterating the shooter's position. The model was trained on a real-world gunshot dataset and optimized for deployment on the ultra-low-power STM32U5 microcontroller. A series of incremental experiments on architectural design, data manipulation, and feature selection aimed to improve performance and efficiency.

To evaluate detection performance under class imbalance, independent of confidence thresholds, a novel onset-based area under the precision–recall curve (AUPRC) metric was proposed. Computational cost was evaluated through STM32U5 inference benchmarks. Experimental results showed that time-shift augmentation provided the largest performance gain, followed by modest improvements from regularization techniques. Class rebalancing and background noise augmentation had minor effects. Replacing standard convolutions with DSConv substantially improved efficiency. Finally, using ∆mel-frequency cepstral coefficients (∆MFCCs) as input features further improved both performance and efficiency.

Overall, the final ∆MFCC DSConv-GRU model outperformed the quasi-DenseNet baseline in F1-score (+2.6%) while reducing multiply-accumulate operations (-96.9%), RAM usage (-95.8%), and runtime (-97.3%). These improvements enabled real-time inference on microcontrollers, demonstrating that a lightweight deep learning model can perform effectively under strict resource constraints. Hence, this work provides a foundational step toward future embedded gunshot detection systems for wildlife monitoring and anti-poaching applications.

Microtask crowdsourcing workers, also known as crowd workers, perform small tasks known as microtasks.
These people use crowdsourcing platforms to complete these microtasks.
Crowd workers have to work in front of a screen to complete these microtasks, risking musculoskeletal problems and other mental problems.
Their working conditions look similar to desk workers, who are people that work remotely or at the office behind a desk.
This study aims to find the health differences between crowd workers and desk workers.
It will provide a general overview on the subjective well-being, experienced and mental health.
In order to analyze the differences in health, a survey will be deployed on a crowdsourcing platform in order to recruit crowd workers and desk workers will be recruited through snowball sampling.
The questions of the survey are divided into 5 groups, each representing a health category: general health, workspace quality, physical well-being, social well-being and emotional well-being.
For this study 17 crowd workers were recruited and 9 desk workers.
From the results, desk workers are healthier in general, have a healthier workspace because some desk workers work in ergonomically good offices, a healthier physical well-being, a healthier social well-being due to them having colleagues and a better emotional well-being. Crowd workers have a lower level of stress, because of the microtasks being mostly very simple, while desk workers have mentally demanding deadlines and projects to work on.