| Object detection is one of the basic tasks in the field of computer vision.The existing 2D object detection algorithms can give the rectangular bounding box and the corresponding category of the object in the RGB image.However,2D object detection has limitations and cannot meet the specific needs of people living in a three-dimensional world.3D object detection will give the object’s 3D bounding box,which plays an important role in certain application scenarios such as autonomous driving,robotics,and 3D scene reconstruction.Compared with other sensors,the monocular camera has the characteristics of low price,easy acquisition,so it has potential commercial value and far-reaching research significance to complete 3D object detection based on monocular vision.The main research content and contributions of this paper are as follows:This paper divides the 3D object detection task into several subtasks,including 2D object detection module,dimension prediction module,orientation prediction module,and geometric constraint module.Each subtask is responsible for estimating specific parameters involved in the 3D object detection task.Decompose a complex problem into several simpler sub-problems,and combine the solutions of the sub-problems to obtain the solution of the original problem.(1)This paper proposes a dimension prediction loss function based on Intersection over Union(IoU).In this paper,the algorithm for calculating the rectangular bounding box IoU is extended to three-dimensional space,and an algorithm for calculating the 3D bounding box IoU in the dimension prediction scene is proposed.Compared with the traditional loss function that calculates the error of each dimensional component of the prediction separately,the loss function based on IoU considers the intrinsic relationship of all the attributes of the bounding box,and has the characteristics of scale invariance.(2)This paper proposes a strategy to indirectly predict the orientation of the object.The realization of 3D object detection based on monocular vision only provides RGB image information,and the orientation is not directly related to the appearance of the object in the RGB image.For this reason,this article chooses to predict the local orientation and calculates the final required global orientation through geometric relations.(3)Aiming at the problem of difficulty in obtaining object depth information when completing 3D object detection task based on monocular vision,this paper uses the principle of geometric constraints to calculate the position coordinates of the object relative to the observer by the least square method.At the same time,this paper proposes an optimized network to modify the object position coordinates.Combining 2D object detection,dimension and orientation prediction results to complete the final 3D object detection task.In the experiment stage,this paper uses multiple indicators to evaluate the performance of the proposed 3D object detection algorithms.The results show that the algorithm proposed in this paper has higher detection accuracy compared with other detection algorithms. |
