Monocular Visual Measurement Based on Marking Points Regression and Semantic Information

This paper studies the visual measurement problem in the intelligent warehouse stocktaking task and designs a set of monocular visual measurement methods that can be used in the UAV-AGV collaborative autonomous warehouse stocktaking system. Monocular visual measurement attempts to complete the measurement task of the target through the information of a single image. Such methods are usually applied to the size estimation of human bodies or small objects and lack practice in large-scale industrial scenes. Compared with the traditional monocular visual measurement algorithm, our method extracts and utilizes the structured information and semantic information in the image. We design the BoxMPR network to undertake the regression task of upper corners in cargo and use LEDNet [1] to identify the semantic information of shelves. Based on the understanding of the scene in an image, we join the real size of the shelf as prior information to measure the target. We systematically evaluate the proposed BoxMPR neural network and our measurement method, which can realize the centimeter-level measurement of the target area in the warehouse scene.