Research on semantic segmentation based on improved BiSeNet on unstructured road

Semantic segmentation is very important for unmanned vehicles on unstructured roads. While accuracy of current algorithms is not high enough especially long distance or tiny objects identification and classification which can lead serious accidents. And the parameters used in calculation are so more as well. This paper proposes a semantic segmentation algorithm based on the BiSeNet model. A lightweight network model MobileNetV3-large is employed as the feature extraction network. Additionally, an RFB module is introduced to increase the receptive field and obtain more global information. The CBAM-FFM module is utilized to combine spatial and channel attention mechanisms feature fusion as well. Experimental results show that the improved model achieves an accuracy of 97.47% on the IDD dataset while reducing the model parameters to 3.05×106. Moreover, it exhibits certain generalization performance on a self-made dataset. Compared to other current semantic segmentation models, the proposed model shows enhanced accuracy, thereby providing important technical support for the practical application of unmanned vehicle on complex unstructured roads.