Sidewalk Pavement Defect Detection with Instance Segmentation: A Step Towards Improving Sidewalk Quality

Making cities walkable and supportive of non-motorized transportation enhances sustainability and inclusivity. Sidewalks, the primary infrastructure in direct contact with pedestrians to support their daily travel needs, deteriorate and develop faults with time and continuous use. However, manual and commonly available non-destructive inspections often are slow, labor-intensive, and cost-inhibitive. This study proposes a quick and cheap method using instance segmentation techniques to detect sidewalk pavement surface defects automatically. The proposed methods are two instance innovative approaches to train the Mask R-CNN (MRCNN) instance segmentation model to identify and provide the respective boundaries of each defect. The first and most reliable of the two is integrating the Simple Copy-Paste Augmentation mechanism for training the MRCNN model. The second mechanism is training via the Cascade method. The results demonstrate the ability of instance segmentation to give sidewalk pavement management teams sufficient information regarding the location, shape, size, and type of damage to make the necessary rehabilitation decision. Inspection can be quick, remote, and frequented, enhancing the sidewalk pavement's resilience and serviceability.