3D Object Detection Based On Deep Learning And Geometric Constraints

Monocular 3D target detection is an important subject in the field of autonomous driving.Existing monocular 3D object detection algorithms are usually based on complete deep learning method.At the same time,it is often only trained for categories with a large number of training samples,such as cars,and other categories with a small sample size,such as cyclists,are not given much attention.For networks that need to complete spatial inference,the number of samples has played a significant role in successfully training the network.Even if it is expanded by means such as data augmentation,the network often has little effect in reasoning about 3D results.In actual autonomous driving scenarios,the correct detection of these categories is also crucial.On the other hand,most algorithms hope that the network can implicitly learn the mapping relationship between two-dimensional pictures and three-dimensional objects without paying much attention to the imaging constraints.However,the explicit application of geometric constraints is also an important idea for 3D target detection.Therefore,this paper focuses on the categories with a small sample size in the autonomous driving scene.By applying the geometric constraints of 3D objects and 2D pictures,a unique multi-stage 3D object detection scheme is proposed.The method mainly includes two stages.In the first stage,a 2D object detection algorithm is used to obtain the 2D bounding box of the object,and in the second stage,the 3D bounding box is derived based on the 2D results.In this paper,the properties of the 3D bounding box derived in the second stage are split into two steps.The first step is to use the visual retrieval method to return the dimensions and orientation of the 3D bounding box.And the second step is based on the 2D results and 3D attributes to obtain the coordinates of the 3D bounding box through the stereo vision geometric constraints.Finally,the two parts are combined to obtain a complete 3D bounding box representation.In the second step of the second stage,this article actually uses three ways to derive 3D results.By constructing geometric constraints,the imaging relationship is applied explicitly,and the derivation of 3D results is directly completed,which greatly reduces the dependence on the amount of sample data and is very suitable for categories with small sample sizes.Finally,we validate the proposed scheme on the KITTI dataset.