采用KPCA和BP神经网络的单目车载红外图像深度估计

提出一种基于监督学习得到深度估计模型的单目车载红外图像深度估计方法。首先用核主成分分析法（KPCA）筛选红外图像特征。将最初提取的红外图像特征用核函数非线性映射到一个线性可分的高维特征空间，再完成主成分分析（PCA），得到降维后的红外图像特征。然后以BP神经网络为模型基础，对红外图像特征和深度值进行训练，训练后的深度估计模型可对单目车载红外图像的深度分布进行估计。实验结果证明，利用该模型估计的单目车载红外图像的深度信息与原红外图像的深度信息一致。

Depth estimation from monocular vehicle infrared images based on KPCA and BP neural network

A depth estimation algorithm from monocular vehicle infrared image based on depth estimation model by supervised learning was proposed. Firstly, kernel-based principle component analysis(KPCA) was used to select infrared image features. Original features extracted from infrared image were project nonlinearly to a high dimensional and linear separable feature space using kernel function. Principle component analysis(PCA) was performed to get dimension reduction infrared image features. Then the infrared image features and depth values were trained using BP neural network. A depth estimation model was obtained which can estimate the depth distribution of monocular vehicle infrared image. The experimental results show that most of the depth estimated by the model is consistent with the original depth information of infrared image.