Street mapping, object detection and traffic load assessment from SVIs and field data using YOLOv3, machine learning and GIS

Street View Imageries (SVIs) facilitates street mapping, extraction and examination of elements from human’s perspective to aid for navigation and better understanding the real time loads on streets. We have selected an unplanned, densely populated residential place with 5 to 6 floor buildings in Batla House, New Delhi where street-level mapping and object detection using Google Street View Application (GSVA) is of high significance. We mapped 82 streets, spanning 14.2 km, and published the videos on GSVA. The objects were detected and classified with an accuracy of 94.2% and 91.4% respectively by YOLOv3 predefined model through ARC-GIS API for Python from the series of images. Our analysis show satisfactory association of human (R2 = 0.77), motor bike (R2 = 0.45) and cars/auto rickshaw (R2 = 0.58) with street width during afternoon hour. The SVIs and real-time objects detection using ML, GIS and field-assisted data combinatorial approach is more feasible for streetscape in unplanned settlements.