Rapid three-dimensional target detection algorithm based on binocular vision

Abstract

The invention provides a rapid three-dimensional target detection algorithm based on binocular vision, and the algorithm selects ResNet-50 as a feature extraction backbone network, replaces all convolution kernels in the ResNet-50 with depth separable convolution to reduce the calculation amount, and then completes the feature extraction task through the fusion of a feature pyramid structure and multi-scale information. Optimal region candidate boxes are selected from the obtained left and right feature maps through a three-dimensional region suggestion network, and the network classification and regression imbalance problem is improved through a unified dynamic sample weighting strategy; returning the left and right region candidate frames to the required size through a region alignment layer; key point prediction is carried out on the left eye candidate frame after regression, prediction of four corner points of the target is completed, and rough three-dimensional information of the target is obtained in combination with right eye information; and the obtained disparity map is further optimized by using the luminosity consistency constraint, so that the quality of the disparity map is improved to obtain higher-precision disparity information. The method has the advantages that a rapid target detection algorithm based on binocular vision is provided, a three-dimensional target detection task can be rapidly completed while certain precision is guaranteed, and the detection speed is increased by about 2.5 times.

Description

technical field [0001] A fast 3D object detection algorithm based on binocular vision. Background technique [0002] In practical application scenarios, 3D target detection can not only complete the pixel coordinate regression and positioning of 2D target detection, but also measure the depth, size and other information of the real world. The acquisition of depth information further promotes automation, driverless and The development of the Internet of Things and other fields has high practical application value. [0003] The 3D target detection algorithm based on binocular vision has the advantages of high detection accuracy, low cost, and little influence from external interference factors. The actual application value and market far exceed similar algorithms, and the detection of target spatial information can further promote automation. development and progress in the field. Contents of the invention [0004] The purpose of the present invention is to solve the probl...

AI Technical Summary

Problems solved by technology

[0004] The purpose of the present invention is to solve the problem of slow detection speed caused by the increase of the detection accuracy of the three-dimensional target detection algorithm based on binocular vision

Images