This thesis investigates detecting potential parkour spots from urban street level imagery. It shows that by using transfer learning it is possible to re-train a convolutional neural network originally trained for general object detection to instead perform the task of parkour spot detection using just a few thousand street level images annotated by parkour hobbyists. The work demonstrates that this machine learning process can be facilitated by street level images easily available through online services such as Google Street View, open-source machine learning frameworks, and publicly available pre-trained neural networks. As such, these methods are widely available even to people without extensive machine learning knowledge. Combining the low training data requirement and readily available tools means that it is possible to deploy machine learning solutions for tasks like parkour spot detection in just a few days. This provides a novel tool for discovering and understanding physical activity opportunities in one's everyday environment, which is useful for researchers and practitioners of fields such as urban design and exercise video games.