Research On 3D Target Detection And Multi-Target Tracking Algorithm Based On Monocular Vision

The importance of computer vision tasks in today’s society is gradually increasing.In particular,3D object detection and multi-object tracking in computer vision have important application value in intelligent transportation,automatic driving and other fields.However,the complexity of the task in 3D target detection and the variability of the real scene where multi-target tracking is located often increase the difficulty of dealing with the problem,resulting in an unsatisfactory actual effect of the algorithm.This paper conducts in-depth research on the possible problems and difficulties in the field of 3D target detection and multi-target tracking,and proposes corresponding improved algorithms and feasible solutions.The main research contents and contributions of this paper are:(1)This paper proposes a position estimation correction algorithm for 3D target detection:This paper mainly discusses the existing problems in the existing position estimation algorithms based on geometric constraints,and proposes an optimization-based position estimation correction algorithm.The proposed correction algorithm can make further adjustments on the basis of the original estimation algorithm,so that the final 3D bounding box and 2D bounding box have better coincidence effect,so that the final position estimation is more accurate.Experiments show the effectiveness of the position estimation correction algorithm proposed in this paper.(2)This paper proposes a graph network model for multi-target tracking:This paper designs a specific graph neural network model for the characteristics of multi-target tracking,including the graph neural network of the target appearance feature and the graph neural network of the motion feature.Under the premise of not losing the tracking accuracy,the network structure designed in this paper is reduced as much as possible,and only the important part is kept,so as to ensure that the network has a relatively ideal operating efficiency and high scalability.Experiments show that the graph network model proposed in this paper has a good effect on multi-target tracking tasks.(3)This paper proposes a joint data association based on Greenkhorn algorithm and KM algorithm:For the data association problem in multi-target tracking,this paper proposes a joint data association algorithm of Greenkhorn algorithm and KM algorithm.The application of the Greenkhorn algorithm is not limited by the pre-module(referring to the feature extraction part)of the tracking algorithm.The similarity matrix of graph matching can be converted into a probability matrix suitable for multi-target tracking,which can be expressed in multi-target tracking tasks.clear physical meaning.Experiments show the effectiveness of the joint data association of Greenkhorn algorithm and KM algorithm.