Three-dimensional Attitude Angle Measurement Of Objects Based On Convolutional Neural Network

Accurate acquisition of the three-dimensional attitude angle of a particular object is particularly important.The effective application of relevant information provides critical information reference and assistance in related fields including industrial production,aerospace,intelligent robots and the like.Traditional 3D attitude angle measurement methods often require high-precision measuring instruments for analysis.The relevant measurement schemes have high application cost and limited applicability.With the development of science and technology and the wide application in the field of computer vision,the three-dimensional attitude angle measurement based on visual analysis has achieved relatively good measurement results.However,the existing three-dimensional attitude angle measurement based on the vision system needs to solve the complex visual correspondence between the feature points on the actual object and the acquired image,and the solution relationship of different feature identifiers is often different.With the rapid development of neural networks,related theories have been widely used in various scenarios such as target detection,motion recognition,behavior analysis,and object classification based on computer vision.Relying on effective visual feature extraction algorithm and network fitting ability,its performance and effect are better than traditional computer vision detection performance and higher precision.Therefore,three-dimensional attitude angle measurement of objects based on this theory has become possible.This paper proposes a three-dimensional attitude angle measurement scheme based on Convolutional Neural Network(CNN),which successfully implements end-to-end detection.Further,in order to effectively expand the measurement range and reduce the amount of neural network training samples.We demonstrate the independence of the change of the information of each attitude angle dimension of the object,and propose the Sub-Added training scheme.At the same time,in order to solve the correspondence between the difference sum image and the original sample image,we propose the Generating Image Encoder to ensure the detection effect of the training model in practical applications.Experiments show that the proposed three-dimensional attitude angle measurement scheme based on convolutional neural network is significantly better than the most advanced method.At the same time,our proposed attitude measurement scheme overcomes the shortcomings of the traditional scheme,and achieves a significant improvement in measurement speed and efficiency,which can achieve the purpose of real-time measurement.