Traffic-sign Recognition from Street-level Photos: a Deep Learning Approach

Traffic-sign Recognition from Street-level Photos: a Deep Learning Approach

The scope of this thesis is the recognition of traffic-signs from street-level images. A state-of-the-art deep learning detection algorithm is used, the Single Shot Multi-box detector (SSD) and within the thesis its performance is validated experimentally by examining different training strategies, modifying the structure of the model and tuning the default bounding boxes as well as other hyperparameters. The final trained model achieves recognition of two kinds of traffic signs: the ‘stop’ and the ‘pedestrian crossing’ sign, with a mean average precision up to 85.90% mAP, testing on a subset randomly chosen from the LISA dataset. Moreover, it is discussed how such a model can be adapted for recognizing traffic signs from crowd-source imagery platforms, such as Mapillary, where images have different properties (e.g. resolution, size and type) and often exhibit barrel distortion.