| Disparity estimation accuracy and real-time performance are two important indicators to evaluate the reliability and practicality of stereo matching algorithm.In the advanced driving assistance system（ADSA）,the imaging quality of outdoor scene images collected by vehicle-mounted binocular cameras is easily affected by the driving road environment,and the accuracy of binocular disparity estimation will be reduced by weak light or complex light environment.At the same time,the safe and reliable driving assistance system requires that the disparity estimation algorithm can accurately and quickly detect the driving environment,and quickly restore the real 3D scene through the 2D data collected by the binocular cameras.Therefore,to ensure the real-time and accuracy of binocular disparity estimation is of great significance to ensure the passive and active safety of ADAS.In order to further improve the real-time performance and accuracy of disparity estimation algorithm,an sparse stereo matching network with attentional mechanism and feature fusion method is studied and experimental research is carried out.The main contents include:（1）Based on the stereo matching network framework,a deep stereo matching network with disparity estimation accuracy and real-time performance is designed.ResNet network with high robustness was used as the backbone feature extraction module,and atrous convolution was used to enhance the high-dimensional feature expression.The cost aggregation of stereo matching is realized by using the 3d line cost aggregation method which improved the performance of disparity estimation.Finally,the disparity estimation feature is calculated by 3d convolution and bilinear interpolation method and then the disparity map is generated by regression.The disparity estimation error and frame rate on the ApolloScape test set are 1.846 px and 13.72 fps,and the frame rate is 64.9% higher than that of PSMNet.Experiment results show that this network can have both estimation accuracy and real-time performance well.（2）Feature extraction and cost aggregation of stereo matching were improved by replacing the prototype backbone network with sparse convolution,and the overall performance of the prototype network was improved by sparse convolution and submanifold sparse convolution.The estimation error and frame rate of the improved network on the test set are 1.599 px and 15.46 fps,and the overall performance is better than that of GA-Net,which further improves the estimation accuracy and real-time performance of the prototype network.（3）An attentional mechanism suitable for sparse features is constructed by improving the computational process of attentional mechanism,and the feature fusion of foreground semantic features and attentional sparse features is realized based on the adaptive linear superposition method.The cost aggregation process of sparse stereo matching network is optimized by computing dense features and sparse features to generate aggregation features with strong data correlation.The experimental results show that the estimation error and frame rate of the attentional feature fusion sparse stereo matching network on the test set are1.498 px and 15.72 fps,and the overall performance is better than that of the improved sparse network.Compared with PSMNet and GA-Net,the estimation accuracy and real-time performance are improved by 13.1% and 9.4%,89.0% and 631.1%,respectively.The performance of disparity estimation is further improved. |
