基于多尺度卷积神经网络的立体匹配算法研究

Research on Stereo Matching Algorithm Based on Multiscale Convolutional Neural Network

摘要:

针对传统障碍物检测中的立体匹配算法存在特征提取不充分，在复杂场景和光照变化明显等区域存在误匹配率较高，算法所获视差图精度较低等问题，提出了一种基于多尺度卷积神经网络的立体匹配方法。首先，在匹配代价计算阶段，建立了一种基于多尺度卷积神经网络模型，采用多尺度卷积神经网络捕获图像的多尺度特征。为增强模型的抗干扰和快速收敛能力，在原有损失函数中提出改进，使新的损失函数在训练时可以由一正一负两个样本同时进行训练，缩短了模型训练时间。其次，在代价聚合阶段，构造一个全局能量函数，将二维图像上的最优问题分解为四个方向上的一维问题，利用动态规划的思想，得到最优视差。最后，通过左右一致性检测对所得视差进行进一步精化，得到最终视差图。在Middlebury数据集提供的标准立体匹配图像测试对上进行了对比实验，经过实验验证算法的平均误匹配率为4.94%，小于对比实验结果，并提高了在光照变化明显以及复杂区域的匹配精度，得到了高精度视差图。

Abstract:

Aiming at the problems of insufficient feature extraction in stereo matching algorithms in traditional obstacle detection, high mismatch rates in areas such as complex scenes and obvious lighting changes, and low accuracy of disparity maps obtained by the algorithm, a multi-scale based Stereo matching method of convolutional neural network is proposed. First, in the stage of calculating the matching cost, a multi-scale convolutional neural network model is established, and the multi-scale convolutional neural network is used to capture the multi-scale features of the image. In order to enhance the model"s anti-interference and fast convergence capabilities, improvements are proposed in the original loss function, so that the new loss function can be trained simultaneously with two positive and one negative samples during training, which shortens the model training time. Secondly, in the cost aggregation stage, a global energy function is constructed to decompose the optimal problem on a two-dimensional image into a one-dimensional problem in four directions. Using the idea of ??dynamic programming, the optimal parallax is obtained. Finally, the obtained parallax is further refined through left-right consistency detection to obtain a final parallax map. A comparison experiment was performed on the standard stereo matching image test pair provided by the Middlebury dataset. The average error matching rate of the algorithm verified by the experiment was 4.94%, which is less than the comparison experiment results. The accuracy of matching in obvious illumination changes and complex regions is improved. A high-precision parallax map was obtained through experiment.