深度学习算法在无人驾驶视觉中的应用

发布时间：2024-01-20 07:24

　　近年来,深度学习技术的研究促进了人工智能在学术界和工业界的发展。深度学习算法起源于人工神经网络,为多层神经网络在实际中的应用提供了一种有效的途径。得益于互联网推动下大数据的积累,以及基于图形处理单元(GPU)的并行计算能力的提升,两者正在促进深度学习算法应用到更广泛的领域,如无人驾驶。无人驾驶因能带了更安全的驾驶体验,降低交通事故发生率,并能有效减少城市中的交通拥挤而受到广泛关注。无人驾驶汽车是一个复杂的系统,而视觉感知是无人驾驶中很重要的一个组成部分。无人驾驶中的视觉感知负责理解周围环境中道路、车辆及行人等。在道路检测方面,传统方法多聚焦于结构化道路和单一道路的情况。但当无人驾驶汽车行驶在自然环境中,它通常需要应对更复杂的道路条件,如边界模糊,凹凸不平的道路,有树阴遮挡的道路,有多条道路同时存在的路口情况等。另一方面,传统算法通常采用浅层的特征提取用于物体的检测与识别。而浅层特征难以表达物体本身具有的抽象特征,因此难以应对同类物体中多变的形态,并且难以区分不同物体中相似的特征。另外,无人驾驶中的实时性要求也限制了计算复杂度较高的传统方法在实际场景中的应用。鉴于由多层神经网络计算出来的...

【文章页数】：155 页

【学位级别】：博士

【文章目录】：
Abstract
摘要
List of Abbreviations
Chapter 1 Introduction
    1.1 Background
        1.1.1 Introduction of Self-driving Vehicle
        1.1.2 Visual Perception In Self-driving Vehicle System
        1.1.3 Road Detection
        1.1.4 Object Detection and Recognition
    1.2 Problems and Challenges
        1.2.1 Road Detection
        1.2.2 Object Detection and Recognition
    1.3 Dissertation Outline
Chapter 2 Fundamentals of Deep Learning
    2.1 Introduction
    2.2 Neural Network in Early Stage
        2.2.1 Multilayer Perception
        2.2.2 Restricted Boltzmann Machine
    2.3 Convolutional Neural Network
        2.3.1 Basic Modules in CNN structures
        2.3.2 Typical CNN-Based Models for Object Recognition
        2.3.3 Platforms for CNN Model Development
    2.4 Conclusion
Chapter 3 CNN-based Road-direction Point Detection
    3.1 Introduction
    3.2 Proposed Method
        3.2.1 Road-direction Point Representation
        3.2.2 Design of Road-direction Point Detection Model
        3.2.3 Loss function
        3.2.4 Convolutional Neural Network Structure
        3.2.5 Non-maximum suppression
        3.2.6 Training of This Model
    3.3 Simulation Results
        3.3.1 Design of Dataset About Road
        3.3.2 Model Simulation
        3.3.3 Performance Evaluation
        3.3.4 Performance Comparison
        3.3.5 Runtime Comparison
    3.4 Discussion
    3.5 Conclusion
Chapter 4 CNN-based Multiple Road-Points Detection
    4.1 Introduction
    4.2 Proposed Method
        4.2.1 Road Representation by Road Points
        4.2.2 Design of Road-Points Detection Model
        4.2.3 Loss Function
        4.2.4 Convolutional Neural Network Structure
        4.2.5 Non-maximum Suppression
        4.2.6 Training of This Model
        4.2.7 Metric Definition for Model Performance
    4.3 Simulation Results
        4.3.1 Design of Road Dataset
        4.3.2 Model Simulation
        4.3.3 Error Analysis
        4.3.4 Mean Performance on Different Categories
        4.3.5 Performance Comparison
        4.3.6 Runtime Comparison
    4.4 Conclusion
Chapter 5 Road-direction Point based Car Detection
    5.1 Introduction
    5.2 Proposed Method
        5.2.1 CNN-based Model with Sub-regions
        5.2.2 Information Integration
        5.2.3 Loss Function
        5.2.4 Convolutional Neural Network Structure
        5.2.5 Training of this model
    5.3 Simulatoin Results
        5.3.1 Preparation of Dataset
        5.3.2 Model Simulation
        5.3.3 Model Performance
        5.3.4 Analysis of Model Performance
        5.3.5 Runtime Comparison
    5.4 Conclusion
Chapter 6 Research on Invariant Object Recognition
    6.1 Introduction
    6.2 Proposed Method
        6.2.1 Sparse Deep Belief Network
        6.2.2 V2 Features Detection with 2-Stage DBN
        6.2.3 SOM Model with Trace Rule
        6.2.4 Metric Method for Model Performance
    6.3 Simulation Results
        6.3.1 Design of Simulation
        6.3.2 Influence of Trace Rule on Performance
        6.3.3 Influence of the Number of SOM Layers
        6.3.4 Influence of Random Order on Performance
        6.3.5 Comparison of Firing Rate of SSI-Top neurons
        6.3.6 Application of Learned SOM Layer
    6.4 Discussion
    6.5 Conclusion
Chapter 7 Conclusions and Future Works
    7.1 Conclusions
    7.2 Future Works
Appendix
References
Acknowledgements
Biography



本文编号：3880424