A Lightweight and Efficient Network for Logistics Truck Scene Semantic Segmentation

Abstract

Intelligent video analytics (IVA) plays an increasingly important role in industrial applications such as autonomous driving, defect detection and logistic scene analytics. Semantic segmentation is a basic technique of IVA, as well as lightweight and efficient segmentation is especially significant in industrial applications. But most excellent methods focus on accuracy, rather than lightweight and fast. Therefore, we propose a unique lightweight and efficient semantic segmentation network aimed at logistics truck scene analysis. Firstly, the lightweight high-level semantic feature extractor, which is a deep branch, is proposed to get a larger receptive field with lower parameters. Secondly, the detail block as a shallow branch is designed to enrich the spatial detail information. Lastly, the feature guided aggregation module based on the novel two-branch methods proposed to guide feature fusion for improving the performance. The parameters of our model are only 1.05M, which are much less than other models, such as PSPNet, SegNet and BiseNet. Compared with some typical real-time segmentation methods, our approach achieves more competitive accuracy (83.94%) and higher inference speed (157 FPS) in the logistics truck scene dataset.