A 3D target detection algorithm for detecting vehicles based on binocular vision

In autonomous driving, the detection of 3D targets in vehicles is an important scene understanding task. Compared to expensive radar devices, 3D target detection methods with the aid of binocular devices have the advantage of low cost and accurate localisation.This paper proposes a three-dimensional object detection OC-3DNet algorithm for binocular vision based on the Stereo Region Convolutional Neural Network (Stereo RCNN), which effectively improves the detection accuracy. To solve the contradiction between the high resolution of the feature extraction part of the network and the perceptual field, this paper combines the Attention-guided Feature Pyramid Network (AC-FPN) after the feature extraction network to improve the detection accuracy of the algorithm for small targets.To solve the problem of large errors in 3D centre projection detection, this paper proposes to establish a constraint relationship between the 3D centre projection and the 2D centre, which further improves the accuracy of 3D object detection.Experimental results show that the improved OC-3DNet algorithm has an average accuracy of 43% on 3D target detection with a threshold of 0.7, which is about 3% improvement over the average accuracy of Stereo R-CNN 3D target detection.