An integration system of AI cars detection with enclosed photogrammetry for indoor parking lot

Abstract

Vehicles localization in a complex built indoor parking lot (e.g. underground parking lot, multistorey parking lot) where the GPS signal is hard to be received is one of the critical tasks for establishing Smart parking system. This research deals with designing a vehicles localization system by using stationary cameras around the indoor parking lot based on AI cars recognition technology integrated with close-range photogrammetry. The test field is located at the parking lot behind the writer's yard, and the main experiment objects are dwellers’ cars. The novel system employs two wireless cameras to shoot the real-time parking lot of photos from the different position for cars detection based on YOLOv3 model. The relative distance between cars and cameras is determined by photogrammetry algorithm by building up stereo pairs of specified cars position between two images with the Oriented FAST and Rotated BRIEF (ORB) feature points. The experiment result shows that the Yolov3 performs relatively well from time-costing and precision perspectives. However, the precision of real-site cars localization initialized by enclose-range photogrammetry algorithm is affected by image noise and the lighting condition intensively. Furthermore, due to the lack of on-site control point and Real-time kinematic (RTK) devices, this project does not convert the local coordinate into the geodetic coordinate system, and this needs to be improved in future research. In conclusion, the integration of Convolutional Neural Network and close-range photogrammetry provide an effective solution for cars positioning under an enclosed scenario with a relatively low budget.