【作者】 王勇；

【导师】 王科；

【作者基本信息】 重庆大学 ， 机械（专业学位）， 2024， 硕士

【摘要】 在自动驾驶领域,随着其智能化等级逐步提高,从封闭道路走向开放道路,对车辆定位提出了更高的要求。想要智能汽车完全实现高等级的自动驾驶,就必须依赖于高定位精度和高鲁棒性的定位结果。过去的定位算法主要是基于低层次的几何特征,在面对复杂环境时其性能往往受限。尽管深度学习可以更好的捕捉环境中的语义信息,但它们会忽略几何和结构信息的重要性,以至于基于纯深度学习的定位方法在一些场景中,定位的精度达不到几何方法的水准。针对复杂场景下的车辆定位问题,本文提出的方法可以分为显著性结构化特征的视觉里程计方法以及显著性点线的回环检测方法两个部分:(1)提出的视觉显著性结构化特征的里程计算法,通过提取环境中多种几何特征以及使用显著性方法模仿人类在判断场景相似性的机制。设计改进的图像增强策略解决了低光环境带来的错误匹配的问题。提出的显著性点线特征策略实现了在应对弱纹理导致的点特征不足的同时又可以节约计算资源。提出的使用曼哈顿假设用于局部地图优化中减少相机位姿漂移的方法。(2)本文在显著性点线的回环检测算法部分,提出的显著性的点线的词袋模型和相似度计算方法,可以解决部分场景中因为点特征相似性导致基于特征点词袋模型错误回环的问题。提出的改进k-means++的聚类策略,在使用k-means++计算数据点成为新的质心点时加入了显著性的权重,在保证聚类中心分布的均匀的同时,又可以让期望关注的特征更容易成为聚类中心。本文分别在公开的数据集、复杂环境数据集、和校园户外车库数据集对提出的里程计算法和回环检测算法进行了测试。测试结果说明了:基于显著性结构化特征的视觉里程计算法在公开数据集上和其他先进的算法相比具有更高的鲁棒性和定位精度,在复杂环境和校园场景中依然具有稳定可靠的定位效果,同时本文提出的基于显著性的点线词袋模型相对于传统的词袋模型,可以有效的降低错误回环的发生,提升定位算法的精度和鲁棒性。更多还原

【Abstract】 In the field of autonomous driving,as its level of intelligence gradually increases,from closed roads to open roads,higher requirements are put forward for vehicle positioning.For Intelligent Vehicles to fully realize high-level autonomous driving,they must rely on positioning results with high positioning accuracy and high robustness.Past positioning algorithms were mainly based on low-level geometric features,and their performance was often limited when facing complex environments.Although deep learning can better capture environmental information,they ignore the importance of geometric and structural information,so that in some scenarios,the positioning accuracy of pure deep learning-based positioning methods cannot reach the level of geometric methods.For the vehicle positioning problem in complex scenes,the method proposed in this article can be divided into two parts:the visual odometry method of salient structured features and the loop detection method of salient point lines:(1)A visual saliency-based structured features odometry algorithm is proposed.By extracting various geometric features from the environment and employing saliency methods to mimic the human mechanism of judging scene similarity,the algorithm enhances its performance in complex scenarios.Additionally,an improved image enhancement strategy is designed to address the issue of mismatching in low-light environments.Furthermore,a salient point and line feature strategy is introduced to conserve computing resources while addressing the insufficiency of point features in weak-texture scenes.Lastly,a method utilizing the Manhattan hypothesis is proposed to mitigate camera pose drift in local map optimization.(2)In the loop detection algorithm of salient points and lines,this paper proposes a bag-of-words model and similarity calculation method for salient points and lines,which solves the problem of bag-of-words based on feature points due to similarity of point features in some scenarios.The problem of model error loopback.A clustering strategy that improves k-means++is proposed.When k-means++is used to calculate data points to become new centroid points,significance weights are added.This ensures the uniform distribution of cluster centers and makes it worthy of our attention.Features are more likely to become cluster centers.This paper tested the proposed mileage calculation method and loop detection algorithm on public data sets,complex environment data sets,and campus outdoor garage data sets.The test results show that the visual odometry calculation method based on salient structural features has higher robustness and positioning accuracy than other advanced algorithms on public data sets,and is still stable in complex environments and campus scenes.Reliable positioning effect.At the same time,the point-line bag-of-words model based on saliency proposed in this article can effectively reduce the occurrence of error loops and improve the accuracy and robustness of the positioning algorithm compared with the traditional bag-of-words model.更多还原