Improved Disparity Estimation Algorithm Based on PSMNet

When designing a network for disparity prediction, increasing the number of convolutional layers can effectively improve the accuracy. However, a too deep network is likely to cause too many parameters and long running time, which is not conducive to real-time disparity estimation. To solve this problem, an improved disparity estimation algorithm based on PSMNet is proposed. First, the dimensionality of the input image is adjusted through a simple convolutional layer. The multi-level residual network structure is used for feature information extraction, and rich image feature information of left and right images is obtained. Then the SPP structure is used to extract and fuse image features at multiple scales. Finally, the concatenated left and right image features are input as a cost volume to the 3D convolutional layer for disparity regression calculation, and the predicted disparity is obtained. The experimental results show that our algorithm reduce the number of parameters by 46.6% compared to the PSMNet. At the same time, the accuracy of the model in the local environment has increased by 7%. This algorithm reduces the number of model parameters while ensuring the accuracy of disparity estimation.