【作者】 王博；

【导师】 吴忻生； 谢宏威；

【作者基本信息】 华南理工大学 ， 控制工程（专业学位）， 2021， 硕士

【摘要】 地点识别是机器人同步定位与地图构建技术中非常重要的环节。随着视觉识别任务的复杂程度越来越高,传统地点识别算法识别效果差的缺点越来越突出。准确快速的地点识别算法可以提高机器人定位建图的鲁棒性。近年来深度学习方法取得了众多成果,在应对大范围的光照变化以及图像中对象的遮挡情况有更好的表现。本文首先介绍了传统地点识别算法,然后介绍了基于深度学习的地点识别算法流程,对比了传统地点识别方法和深度学习方法的优缺点,提出了基于卷积神经网络的提高精确度和提高推理速度的地点识别算法模型。(1).针对已有地点识别算法精度不高的问题,设计了融合深层语义特征的地点识别算法。首先设计深层卷积模块提取图像更丰富的语义信息,然后将卷积特征送入局部聚合向量模块生成地点描述符,最后在数据集训练得到最优模型参数。使用东京街景数据集和匹兹堡数据集训练模型,验证了本方案在对应测试集上召回率优于对比算法。在通用地点检索数据集上测试所得模型的泛化性能,测试结果表明了本方案所得模型的泛化能力优于对比算法。(2).考虑到地点识别算法部署的实时性,设计了融合注意力模块的轻量化地点识别算法。首先将特征提取网络替换为轻量化网络结构,然后融合高效通道注意力算法得到图像特征提取模块,最后将提取到的图像特征送入局部聚合向量模块得到地点描述符,在匹兹堡数据集上训练得到最优模型参数。通过对比现有地点识别算法,验证了本方案在保证精度的情况下压缩了模型参数,明显提升了地点识别算法的推理速度,显著减少了模型的参数量,实验表明算法可以在实时性要求较高的场景下使用。更多还原

【Abstract】 Place recognition is a very important part of robot simultaneous localization and mapping technology.With the increasing complexity of visual recognition task,the disadvantage of poor recognition result of traditional place recognition algorithm is more and more prominent.Accurate and fast place recognition algorithm can improve the robustness of robot localization and mapping.In recent years,the deep learning scheme has achieved a lot of results,which has a better performance in dealing with a wide range of illumination changes and occlusion of objects in the image.This paper first introduces the traditional place recognition algorithm,and then introduces the algorithm process based on deep learning,compares the advantages and disadvantages of traditional place recognition method and deep learning method,and proposes palce recognition algorithm models based on convolution neural network to improve the accuracy and prediction speed.(1).Aiming at the problem of low accuracy of existing place recognition algorithms,a place recognition algorithm based on deep semantic features is designed.Firstly,the deep convolution module is designed to extract more semantic information of the image,then the convolution features are sent to the local aggregation vector module to generate the place descriptor,and finally the optimal model parameters are obtained by training on the datasets.Using the Tokyo street view dataset and Pittsburgh dataset training model,it is verified that the recall rate of this algorithm is better than the comparison algorithm in the corresponding testset.The generalization performance of the model is tested on the general place retrieval dataset,and the test results show that the generalization ability of the model is better than that of the comparison algorithm.(2).Considering the real-time deployment of place recognition algorithm,a lightweight place recognition algorithm based on attention model is designed.Firstly,the feature extraction network is replaced by lightweight network structure,then the image feature extraction module is obtained by combining the efficiency channel attention algorithm.Finally,the extracted image features are sent to the local aggregation vector module to get the place descriptor,and the optimal model parameters are trained on the Pittsburgh dataset.By comparing with the existing place recognition algorithm,it is verified that this scheme can compress the model parameters while ensuring the accuracy.The proposed lightweight place recognition scheme significantly improve the prediction speed of the place recognition algorithm,and also significantly reduce the amount of model parameters.Experiments show that the proposed lightweight place recognition algorithm based on attention model can be used in the scene with high real-time requirements.更多还原