【作者】 吴琨；

【导师】 刘琼； 李繁；

【作者基本信息】 华南理工大学 ， 工程硕士（专业学位）， 2020， 硕士

【摘要】 车载热成像行人检测系统通过热成像传感器获取场景信息,使用模式识别技术检测道路场景中的行人。现有的方法多将图像行人与背景作为二类分类问题处理。道路场景动态,行人属非刚体,外观多变,行人样本与背景样本的分类边界呈非线性,线性分类器很难准确分类道路场景中的行人和背景。本文通过分析数据分布,采用子类划分减小行人类内差异,依此优化数据分布和分类器结构,提高车载热成像行人检测模型性能。1.传感器视角和行人姿态连续变化,道路场景行人多样性显著,样本类内差异较大。依据行人特征进行子类划分,人工子类划分效率效果受限,采用聚类算法自动划分行人子类。高维数据聚类难以收敛,采用流形学习在保留行人结构特征的前提下将高维行人特征映射为低维表达。2.实验表明各个子类样本数量不平衡影响深度卷积网络模型性能,针对热成像行人设计数据平衡方法,促进模型学习不同行人外观的共性特征。采用重采样使行人子类样本数量趋于平衡;为防止重采样后模型过拟合,设计热成像图像实例切换策略,增强行人样本多样性;优化损失函数调整各行人子类学习权重。实验表明经优化处理,模型的检测精度得到提高。3.系统级集成优化。通过子类训练获得多个基分类器,根据分类器权值整合基分类器得分构建级联分类器。改进RoIs筛选机制,采用得分衰减和局部滑窗机制,降低漏检率,减少虚警。实验表明级联分类器与核函数SVM性能接近而运算量更低,适合DSP等嵌入式应用。更多还原

【Abstract】 Vehicle-mounted thermal imaging pedestrian detection system obtains scene information through thermal imaging sensor and use pattern recognition technology to detect pedestrians in road scenes.Most of the existing methods treat image pedestrian and background as two-class classification problems.The road scene is dynamic.Pedestrians are non-rigid and their appearance is changeable.The classification boundary of pedestrian samples and background samples is nonlinear,so it is difficult for linear classifier to accurately classify pedestrians and background in the road scene.In this paper,the data distribution is analyzed,and subclasses are adopted to reduce the intra-row differences.Based on this,the data distribution and classifier structure are optimized to improve the performance of vehicle-mounted thermal imaging pedestrian detection model.1.The sensor angle and pedestrian attitude change continuously,the pedestrian diversity in the road scene is significant,and the difference within the sample class is large.Based on pedestrian characteristics,pedestrian subclasses are subclassed,and the efficiency of artificial subclasses is limited.The clustering of high-dimensional data is difficult to converge,so manifold learning is used to map high-dimensional pedestrian characteristics to low-dimensional expressions on the premise of preserving pedestrian structural characteristics.2.The experiment shows that the imbalance of samples of each subclass affects the performance of the deep convolutional network model,and the data balance method is designed for the thermal imaging pedestrian to promote the model to learn the common characteristics of different pedestrian appearances.Resampling is used to balance the number of pedestrian subclasses.In order to prevent the over-fitting of the model after resampling,the case switching strategy of thermal imaging image was designed to enhance the diversity of pedestrian samples.The optimization loss function adjusts the learning weight of each pedestrian subclass.The experimental results show that the detection accuracy of the model is improved.3.System-level integration optimization.Multiple base classifiers were obtained by subclass training,and a cascade classifier was constructed by integrating base classifier scores according to classifier weights.Improve RoIs screening mechanism,adopt score attenuation and local sliding window mechanism,reduce the rate of missed detection,reduce false alarm.Experiments show that the performance of cascade classifier is close to that of kernel function SVM and the computation is lower,which is suitable for embedded applications such as DSP.更多还原