摘要
盲人作为弱势群体,他们的衣食住行值得被人们关注,其中出行问题是造成盲人群体远离社会的重要原因。本文提出了一种基于DenseNet改进的深度估计算法,以解决盲人出行时无法感知周围障碍物的问题。首先,以DenseNet作为编码器的编解码过程中,信息丢失会造成深度估计不准确,为了减少这种问题在编码器与解码器的跳跃连接中引入RHAG残差混合注意力组,加强模型对细节特征的识别能力,提升模型恢复深度信息的准确性;然后,在解码出深度图后采用AdaBins后处理模块,对深度图进行优化,以更好地恢复出RGB场景的深度信息;最后通过ACB非对称卷积替换DenseNet中DenseBlock的卷积,通过增强卷积骨架,提升模型特征提取能力。实验结果表明,本文改进的算法与原网络相比,精度提升了约3.04%,均方根误差降低了约3.39%。与目前先进的深度估计网络MonoDepth相比,精度提升了约2.2%,绝对相对误差降低了约1.3%。本文算法在通过单张RGB图进行深度估计时能获取到更准确的深度信息,优于对比算法,且满足边缘计算设备的要求,具有一定的实用价值。
Blind people, as a vulnerable group, deserve attention for their clothing, food, housing, and trans-portation, among which transportation issues are an important reason for the blind group to stay away from society. This article proposes an improved depth estimation algorithm based on Dense-Net to solve the problem of blind people being unable to perceive surrounding obstacles when trav-eling. Firstly, information loss during the encoding and decoding process can lead to inaccurate depth estimation. In order to reduce this problem, RHAG residual mixed attention groups are in-troduced in the skip connection between the encoder and decoder to enhance the model’s ability to recognize detailed features and improve the accuracy of the model in restoring depth information;then, the AdaBins post-processing module is used to optimize the depth map to better recover the depth information of RGB scenes;finally, ACB asymmetric convolution is used to replace the convo-lution of DenseBlock in DenseNet, enhancing the model’s feature extraction ability by enhancing the convolution skeleton. The experimental results show that the accuracy of the improved algorithm is improved by 3.04% and the root-mean-square deviation is reduced by 3.39% compared with the original network. Compared with the current advanced depth estimation network MonoDepth, the accuracy has been improved by about 2.2% and the absolute relative error has been reduced by about 1.3%. The algorithm in this paper can obtain more accurate depth information when depth estimation is carried out through a single RGB image, which is superior to the comparison algorithm and meets the requirements of edge computing devices, and has certain practical value.
出处
《建模与仿真》
2023年第5期4642-4653,共12页
Modeling and Simulation
