面向路面预瞄的立体匹配网络性能提升策略OA

To solve the problems of the stereo matching networks application to road preview systems with insufficient training data,performance degradation of real-time network and poor stability for disparity prediction,the performance enhancement strategy of stereo matching network for road preview was investigated.The large-scale synthetic dataset tailored for stereo matching was rendered using Unreal Engine 5,and the effectiveness in boosting network performance was verified.Leveraging the theory of distillation learning,the high-performance stereo matching network was employed to generate pseudo ground-truth disparities for manually collected real-world data,and the disparities were used to train the real-time stereo matching network.During the training stage,the disparity consistency loss supervision term was introduced into the loss function to enhance the stability of disparity prediction.The test results on the RSRD dataset show that the DLNR network finetuned with the road preview synthetic dataset achieves 84.5％reduction in the end-point error(EPE)metric.On the locally collected road-preview dataset,the Fast-ACVNet network trained via distillation learning exhibits 65.6％decrease in EPE,while the introduction of the disparity consistency loss leads to 55.9％improvement in disparity prediction stability.In speed-bump distance measurement trials,the Fast-ACVNet network trained by the proposed comprehensive strategies shows the ranging error within 30 cm,indicating outstanding accuracy and stability with favorable road preview performance.