Method for detecting three-dimensional object based on point fusion network

Abstract

The invention puts forward a method for detecting a three-dimensional object based on a point fusion network. The method comprises a point cloud network, a fusion network, and dense fusion predictionscoring functions. The method comprises the following steps: a point cloud network model is used for taking in original point cloud, learning spatial coding of each point and aggregating global pointcloud features; the features are used for classification and semantic segmentation; in the fusion network, image features extracted via a convolutional neural network and corresponding point cloud features generated via a sub-network of the point fusion network are used as input; these functions are combined, a three-dimensional bounding box is output for a target object, a supervisory scoring function is used to directly train the network to predict whether points are within the target bounding box, and an unsupervised scoring function can help the network select a best prediction point. Viathe method, how to combine images and depth information optimally can be directly learned; lossy input preprocessing such as quantization or projection or the like can be prevented; the method can begenerally applicable, and accuracy thereof is greatly improved.

Description

technical field [0001] The invention relates to the field of object detection, in particular to a three-dimensional object detection method based on a point fusion network. Background technique [0002] The detection and recognition of 3D objects is an important research direction in the field of computer vision, and point fusion is a general 3D object detection method, which can use both image and 3D point cloud information. The application of 3D point cloud data in industry, especially in reverse engineering, has become more and more popular. The main application scenario of 3D object detection and recognition on point cloud data is to identify the existing 3D object model from the obtained point cloud data. 3D object detection technology can be applied to a variety of urban planning, construction and management projects to detect pedestrians, vehicles, shops, etc. in urban scenes; it can also be assisted by automatic detection of pedestrians, vehicles, traffic signs and ...

AI Technical Summary

Problems solved by technology

[0004] Aiming at the problem that existing methods do not have universal applicability, the purpose of the present invention is to provide a 3D object detection method based on point fusion network. The point cloud network model ingests the original point cloud, learns the spatial encoding of each point and aggregates the global points Cloud features, these features are used for classification and semantic segmentation, the fusion network takes the image features extracted by the convolutional neural network and the corresponding point cloud features generated by the sub-network of the point fusion network as input, it combines these functions and generates the target The object outputs a 3D bounding box, and the supervised scoring function is used to directly train the network to predict whether the point is within the target bounding box, while the unsupervised scoring function can help the network choose the best predicted point

Images