To build advanced prediction models for human behaviour, a solid detection of relevant objects using a camera video stream is required. Although state-of-the-art vision and machine learning models were used for object detection, there was a significant drop in performance in adverse conditions, such as during night-time or poor weather. I was responsible for finding ways to improve the performance of the object detection in such adverse conditions and researched and sophisticated data augmentation methods which relied on generative adversarial networks to generate more artificial training data of adverse scenarios. In addition, I built deep learning models that could identify multiple types of adverse conditions, which were integrated in the company’s key-frame extraction pipeline of their internal dataset labelling pipeline. This way, more collected training data from adverse conditions would be labelled, which would allow the object detection models to improve their performance continuously over time.