Combining Deep Learning and Street View Imagery for Urban Safety Analysis : Developing an Object Detection System to Assess Safety Perceptions in Stockholm

Sammanfattning

The urban environment is designed to ensure that the quality of life of the citizens is the best possible. Thus, safety perceptions play an essential role in how urban planning and policy decisions are made. To guarantee that the well-being of the citizens is considered while building settlements, continuous and rigorous studies of how these places evolve need to be carried out. However, said task can often be laborious and time-consuming, requiring exhaustive work from experts within the urban planning field. Following recent trends in artificial intelligence (AI) and deep learning (DL) this project proposes an efficient and scalable approach to address this problem. By applying computer vision technologies to street view imagery and utilizing image analysis techniques, a system can be developed and implemented to identify the factors affecting residents’ sense of security easily. The information provided by the system could assist both experts and local governments in their decision-making processes. In this thesis, an approach different from the commonly used image segmentation techniques is proposed. Two object detection models were custom-trained to detect specific elements that might directly interfere with human safety, such as vehicles, traffic signs, or trees. A You Only Look Once (YOLO) model, well known for its low inference time and high accuracy in terms of object detection, was used as a base. The results, obtained by analyzing street view images of different zones within the city of Stockholm, demonstrate that using these fine-tuned models, achieving mean average precision (mAP) scores of 53.63% and 51.26% respectively, can significantly reduce the time spent by local authorities in observing the streets.