Research On 3D Object Detection And Tracking Algorithm Based On Depth Estimation

Object detection and tracking is one of the important research contents in the field of computer vision.It has broad application prospects in the fields of human-computer interaction,automatic driving and intelligent robots.In recent years,2D object detection technology has achieved good research results in public data sets,and has also performed well in many industrial fields.However,the lack of spatial information in 2D object detection,can not meet high-level applications such as automatic driving,robotics and intelligent human-computer interaction.At present,the mainstream object tracking technology mainly adopts the related filtering method,but there are still some shortcomings in the tracking under the object scale change.This paper applies the object depth information to object detection and tracking,is used to solve the above problems,and proposes a 3D object detection algorithm and a multi-scale object tracking algorithm based on depth estimation optimization.The main research contents of this paper are as follows:(1)A single image depth estimation network is proposed.The network uses convolution neural networks and de-convolution neural networks to construct a single image to depth information map to achieve depth estimation of a single image.The binocular parallax image is used as training data input network to estimate the depth information of image scene.The experimental results show that the network can effectively realize the depth estimation of a single image,and it is easier to train and apply than other existing depth estimation networks,because it does not need depth information as training samples.(2)Aiming at the lack of spatial information in 2D object detection algorithm,a 3D object detection algorithm based on 2D object detection algorithm based on the depth estimation network is proposed.Firstly,the position of the object in the image is detected by using the 2D object detection algorithm,and the object depth information is used to calculate the spatial pose.Then,the 3D dense sample frame of the object is generated based on its spatial attitude,and the error of the re-projection frame and the 2D object detection frame is minimized,and obtain the final detection result.The experimental results show that the algorithm can effectively improve the accuracy of 3D object detection.(3)Aiming at the problem that the 2D object tracking algorithm is sensitive to object scale changes,a multi-scale object tracking algorithm based on depth estimation optimization is proposed.The algorithm firstly trains the prediction model according to the object depth value and the object scale change,then uses the object depth information to predict the object scale,and trains the variable scale related filter tracking template to achieve multi-scale object tracking.The experimental results show that the algorithm performs better than the related filter object tracking in the tracking process of object scale changes.