【作者】 王东兴；

【导师】 郭烈；

【作者基本信息】 大连理工大学 ， 车辆工程， 2021， 硕士

【摘要】 近些年随着人工智能技术的推动和5G功能的广泛应用,智能汽车取得了很快的发展。负责理解汽车周围环境的环境感知模块是智能汽车系统中非常关键的一环,高精度的环境感知模块为智能汽车提供了安全的保障。多目标检测及跟踪算法是环境感知模块非常重要的一部分,目前存在计算量较高,无法在嵌入式平台上实时运行的问题。因此,本文基于轻量级网络的设计原则,对多目标检测及跟踪算法进行改进,并最终在嵌入式平台NVIDIA Jetson AGX Xavier上实现实时运行。首先,基于SSD算法设计了两个高效的卷积神经网络模块,分别是上下文增强模块和特征增强模块。并在两个公开数据集上进行了训练和精度验证,分别是MS COCO和Pascal VOC,且在Pascal VOC上对每个模块进行了消融实验。结果显示当图片输入大小为320x320时,在MS COCO测试集上,本文算法的平均检测精度相对于原SSD算法提高了8.5%,在Pascal VOC测试集上,本文算法的平均检测精度相对于原SSD算法提高了3.7%,尤其是小目标类别Bottle的准确率提升了13.2%。其次,基于滴滴的D2-City数据集制作多目标检测及跟踪的训练集和测试集,考虑到大多数多目标跟踪算法只能在服务器上实时运行的问题,基于Fair MOT算法设计轻量级的多目标跟踪模型。采用了轻量级的骨干网络Shuffle Netv2进行特征提取,然后对上下文增强模块和特征增强模块进行进一步优化并插入到多目标检测及跟踪算法当中,并在三个典型场景进行了精度测试,其中城市道路场景跟踪结果显示当模型中仅加入优化后的上下文增强模块时,MOTA值从51.9%提升到了61.1%,当模型同时加入优化后的上下文增强模块和优化后的特征增强模块时,MOTA值从51.9%提升到了63.2%,且ID Switch值也得到了降低,并且当目标暂时消失而又重新出现时,算法具有重新找回轨迹标签的能力。最终,将整个多目标跟踪算法部署到嵌入式平台NVIDIA Jetson AGX Xavier上,先将卷积神经网络部分由Pytorch模型先转成onnx格式,再由onnx格式转成ncnn格式,然后将算法的后处理部分和跟踪算法由python程序转成C++程序,在嵌入式平台上的时间测试结果显示本文算法处理一张图片耗时33.504 ms,满足实时性要求。综上所述,本文对多目标检测及跟踪算法的精度与运行速度进行了优化,提高了算法的跟踪稳定性,算法部署在嵌入式平台NVIDIA Jetson AGX Xavier上之后,每张图片的处理时间满足实时性要求。更多还原

【Abstract】 In recent years,with the promotion of artificial intelligence technology and the widespread application of 5G functions,intelligent vehicles have achieved rapid development.The environment perception module responsible for understanding the surrounding environment of the car is a very critical part of the smart car system.The high-precision environmental perception module provides a safety guarantee for the smart car.The multi-target detection and tracking algorithm is a very important part of the environment perception module.At present,there is a problem that the computational complexity is high and it cannot run in real time on the embedded platform.Therefore,based on the design principle of lightweight network,this paper improves the multi-target detection and tracking algorithm,and finally realizes real-time operation on the embedded platform NVIDIA Jetson AGX Xavier.First,based on the SSD algorithm,two efficient convolutional neural network modules are designed,namely the context enhancement module and the feature enhancement module.And conducted training and accuracy verification on two public data sets,namely MS COCO and Pascal VOC,and conducted ablation experiments on each module on Pascal VOC.The results show that when the image input size is 320x320,on the MS COCO test set,the average detection accuracy of this algorithm is 8.5% higher than that of the original SSD algorithm.On the Pascal VOC test set,the average detection accuracy of this algorithm is increased by 3.7%compared to the original SSD algorithm,especially the accuracy of the small target category Bottle increased by 13.2%.Secondly,based on Didi’s D2-City data set,a training set and a test set for multi-target detection and tracking are made.Considering the problem that most multi-target tracking algorithms can only run on the server in real time,a lightweight multi-target tracking model is designed based on the Fair MOT algorithm.The lightweight backbone network Shuffle Netv2 is used for feature extraction,and then the context enhancement module and feature enhancement module are further optimized and inserted into the multi-target detection and tracking algorithm,and accuracy tests are carried out in three typical scenarios.Among them,the urban road scene tracking results show that when only the optimized context enhancement module is added to the model,the MOTA value increases from 51.9% to 61.1%.When the model adds the optimized context enhancement module and the optimized feature enhancement module at the same time,the MOTA value increases from 51.9% to 63.2%,and the ID Switch value is also reduced.And when the target temporarily disappears and then reappears,the algorithm has the ability to retrieve the trajectory label.Finally,the entire multi-target tracking algorithm is deployed on the embedded platform NVIDIA Jetson AGX Xavier.First,convert the convolutional neural network part from the Pytorch model to the onnx format,and then convert the onnx format to the ncnn format,and then convert the post-processing part of the algorithm and the tracking algorithm from the python program to the C++ program.The time test results on the embedded platform show that the algorithm in this paper takes 33.504 ms to process a picture,which meets the real-time requirements.In summary,this article optimizes the accuracy and running speed of the multi-target detection and tracking algorithm,and improves the tracking stability of the algorithm.After the algorithm is deployed on the embedded platform NVIDIA Jetson AGX Xavier,the processing time of each image meets the real-time requirements.更多还原