DLNet: A Lightweight Drivable Area Segmentation Method Based on DeepLabV3+

Drivable area segmentation is an important part of ensuring safe travel with intelligent driving. In real traffic scenarios, the drivable region presents different sizes due to obstacle occlusion. In order to take into account the characteristics of multi-scale feature extraction and reduce the computational cost of the model, this paper introduces DLNet, an improved model based on DeepLabV3+, which is good at striking a balance between efficiency and cost.The three improved modules of DLNet, the Network Pruning and Optimization Module of MobileNetV2, the MPAM(Multi-scale pooling of attention module) Module, and the Stacked Upsampling Module, can reduce the number of model parameters and the computational effort of the model, while maintaining a high degree of accuracy. Experimental results on the Cityscapes dataset show that the mean pixel accuracy (MPA) is improved by about 0.6% and the mean intersection over union ratio (MIOU) is improved by 0.4% compared to the original model DeepLabV3+ model. The number of parameters is reduced to only 39.2% of the original model and the computational load is reduced to 12.3%. There is also a large improvement compared to other models.