YOLOv4 for Urban Object Detection: Case of Electronic Inventory in St. Petersburg

The paper presents the results of preparing a labeled dataset from open sources for 11 object classes and the analysis of two well-known object detection methods in the task of urban electronic inventory in Saint Petersburg in Russia under the concept of Smart City methods and technologies. We proposed YOLOv4 for urban object detection such as Windows, Doors, Adv Billboards, Ramps, etc. To do that the first step is data collection from the environment, and data augmentation techniques are employed to generate data. Then the transfer learning method is used to train our dataset with both algorithms YOLOv3 & YOLOv4 and finally the NMS method is used to remove overlaps bounding boxes. To evaluate the performance of both methods RMSE used as a metric. The YOLOv4 method showed better results in object detecting and classifying than YOLOv3 in total and in the context of each class. Based on RMSE metric formula average classification loss after the training model for YOLOv3 is 0.66 and against for YOLOv4 is 0.33. Using YOLOv4 helped us to develop the first version of web-service for automated urban object detection and recognition in real-time that can be scaled and distributed to other districts of the city.