Research On Algorithms Of Planar Object Tracking And Lane Detection Based On DenseNet

Convolutional neural networks have become one of the core technologies in the field of computer vision.In the past ten years,many algorithms have used convolutional neural networks to replace traditional methods to solve various complex vision problems.Driven by massive amounts of data,the convolution operation can extract the features of serialized data in scenarios where little prior knowledge is introduced,and improve the robustness of the model.Planar object tracking and lane detection are both hotspots in computer vision research,and they have been widely used in augmented reality and robot vision fields.This article attempts to use the same end-to-end fully convolutional dense network to solve the above two problems.The main research work is as follows:(1)A planar object tracking algorithm based on a two-stage cascaded thick and thin dense network is proposed.Although traditional planar object tracking algorithms based on keypoints and deep learning can solve tracking tasks in complex scenes,the robustness of their trackers largely depends on the performance of feature descriptors.To solve the problem of how to balance the discriminability and geometric invariance of feature descriptors,the algorithm in this paper firstly uses a two-stage network to learn image feature descriptors in stages,so that the generated descriptors can obtain specific geometric invariance while maintaining the overall discriminant ability.Then,the feature map generated by the network is captured by the weighted hollow space pyramid model to capture dynamic context information,and an image feature descriptor with weighted multi-scale information is generated.This paper also conducts a comparative experiment on the algorithm on the POT benchmark data set,and verifies the overall effective-ness of the algorithm model through comparison with a variety of representative algorithms.Moreover,this article also designed an ablation experiment for each module of the algorithm,and verified the effectiveness of each module.(2)A lane detection algorithm based on fully convolutional dense network and multi-task learning is proposed.In practical applications,lane line detection tasks require extremely high real-time performance,and there are many situations where the visual cues of lane lines are not obvious.For this reason,this paper regards lane line detection as a problem of classifying and selecting images line by line.Compared with the traditional deep learning-based lane line detection method,which regards lane line detection as a semantic segmentation problem,it greatly reduces the problem space and reduces the problem.The time complexity of the algorithm is improved;then in order to improve the feature extraction ability of the model,the algorithm in this paper uses a fully convolution dense network and self-focused distillation to give additional optimization constraints to the feature extraction layer.This paper also designs experiments on the Tu Simple benchmark data set to verify the effectiveness of the algorithm.