【作者】 王新；

【导师】 张宏立；

【作者基本信息】 新疆大学 ， 控制工程（专业学位）， 2021， 硕士

【摘要】 随着机动车保有量的快速增加和道路交通的高速发展,如何保障行车安全,减少交通事故成为人们关心的问题,基于计算机视觉的交通目标识别与检测技术也因此具有重要研究价值。本文针对交通目标的识别与检测开展研究,主要研究工作如下:(1)针对传统机器学习方法存在交通标志识别精度低、鲁棒性差和经典卷积神经网络方法在交通标志图像上识别精度低的问题,本文设计了基于多阶段特征融合的卷积神经网络交通标志识别方法。提出了一种多阶段特征融合的卷积结构,包括迭代深度聚合结构和层次深度聚合结构,有效提升了模型对交通标志图像特征的提取能力。为避免卷积模型出现过拟合,降低交通标志的识别效果,使用数据增强和标签平滑的方法提高模型的泛化能力。实验结果表明本文方法具有精度高、泛化能力强的特点。(2)针对自动驾驶场景下的交通目标检测问题,本文提出了一种迭代聚合的高分辨率网络Anchor-free交通目标检测方法,在Center Net的基础上进行改进。首先引入高分辨率表征骨干网络,高分辨率表征骨干网络可以在特征提取的过程中保持特征图的分辨率,有效地减少图像在下采样过程中损失的空间语义信息。并用迭代聚合的方式对高分辨率表征骨干网络输出的不同分辨率的特征图进行融合,充分利用网络提取的特征信息。最后使用注意力机制优化模型提取的特征信息。实验结果表明本文所提方法具有更高的精度,满足实时检测的性能要求,有良好的鲁棒性。(3)针对目前的卷积模型过大、占用存储过多难以在计算资源有限的平台部署的问题,本文研究了以轻量级卷积神经网络Mobile Net V3为基础的交通目标检测方法。引入Mobile Net V3-Large作为编码器进行特征提取,Mobile Net V3-Large使用的深度可分离卷积可有效降低模型的参数量。分析设计了解码器的网络结构,充分利用了编码器输出的特征,在增加较少的参数量和计算量下进一步提高模型的性能。实验结果表明本文方法在检测精度,模型存储大小,检测速度上达到了一个较好的平衡,是一种有效的轻量级交通目标检测方法。更多还原

【Abstract】 With the rapid increase of the number of motor vehicles and the rapid development of road traffic,how to ensure the safety of driving and reduce traffic accidents has become a concern.Therefore,the traffic target recognition and detection technology based on computer vision has important research value.This paper focuses on the recognition and detection of traffic targets.The main research contents are as follows:(1)In view of the problems of low recognition accuracy and robustness of traffic signs in traditional machine learning method and low recognition accuracy of traffic signs in classical convolution neural network method.This paper designs a multi-stage feature fusion-based convolutional neural network traffic sign recognition method,including iterative deep aggregation structure and hierarchical deep aggregation structure.This method effectively improves the feature extraction ability of traffic sign image,and use data enhancement and label smoothing methods to strengthen the model’s generalization ability.The experimental results show that the method in this paper has the characteristics of high precision and strong generalization ability.(2)Aiming at the problem of traffic target detection in automatic driving scene,this paper proposes an iterative aggregation high-resolution network anchor free traffic target detection method,which is improved on the basis of Center Net.Firstly,the highresolution representation backbone network is introduced.The high-resolution representation backbone network can maintain the resolution of the feature map in the process of feature extraction,and effectively reduce the loss of spatial semantic information in the process of image downsampling.The feature maps of different resolutions output by the high-resolution representation backbone network are fused by iterative aggregation method,making full use of the advantages of network extraction feature information.Finally,attention mechanism is used to optimize the feature information extracted from the model.The experimental results show proposed method has higher accuracy,meets the performance requirements of real-time detection and has good robustness.(3)Aiming at the problem that the current convolution model is too large and occupies too much storage,which is difficult to deploy on the computing platform with limited resources,this paper studies the traffic target detection method based on the lightweight convolution neural network Mobile Netv3.Mobile Netv3-Large is introduced as the encoder to extract features.The depth separable convolution used by Mobile Netv3-Large can effectively reduce the parameters of the model.The network structure of the decoder is analyzed and designed,which makes full use of the output features of the encoder and further improves the performance of the model with less parameters and computation.Experimental results show that the proposed method achieves a good balance in detection accuracy,model storage size and detection speed,and is an effective lightweight traffic target detection method.更多还原