Motion Detection Methods Applied on RGB-D Images for Vehicle Classification on the Edge Computing

Intelligent Transportation Systems rely on data processing methods in real-time conditions. There are many methods to process the data. However, the progress of hardware computational power forces us to reassess the effectiveness of some methods. The first idea for discussion is whether cloud or fog computing is necessary to process computer vision methods in real-time conditions. The second idea for discussion is whether a depth map gained from the 2-D image monocular estimation is suitable as an extra feature to process. Another scientific question is the categorization of the vehicles based on 3-D data by extracted volume or height features. This approach could be useful for time of flight (ToF) camera output. This study simulates the output of a ToF camera with a resolution higher than VGA via the monocular depth map estimation with the convolutional neural network model multiscale depth estimation system. This article proposes the computational architectures for data processing between single-board computers with various computational power. Data processing includes obtaining, feature extraction, and classification via various methods. These methods are image loading, background subtraction, shadow removal, monocular depth estimation, pointcloud calculation, 3-D convex hull, and volume thresholding. Volume thresholding is a reliable approach for categorization into light and heavy vehicles. Background subtractors connected with K-Means are reliable for vehicle detection without shadows.