Research On Road Garbage Detection Algorithm Based On Semantic Segmentation

Realizing the intelligent control of road sweepers is necessary to reduce energy consumption and improve efficiency of urban street cleaning work,and real-time road garbage detection algorithm plays an important part in the intelligent control system for road sweepers.This thesis introduces a real-time road scene garbage detection method based on semantic segmentation to detect and segment garbage of various types under road scene,meanwhile the running speed should meet the real-time requirement of road sweepers in practical application.This thesis focuses on two main difficulties in the task of garbage segmentation under road scene,namely speed and accuracy trade-offs and the high cost of pixel-level labelling.Aiming at coping with the speed and accuracy trade-offs,this thesis proposes a road garbage segmentation model based on improved ERFNet,featured by introducing a modified feature pyramid attention module in the architecture and adopting a composite loss in the training strategy,to improve the result without adding to computational burdens.Besides an additional similarity guidance module is added in the model architecture to guide the segmentation results,so as to reduce the dependence on large number of annotated samples from the perspective of few-shot learning.Aiming at dealing with the high cost of pixel-level labelling,this thesis proposes a labelfree sample augmentation method based on local style transfer,which expands the size and diversity of the training set without requiring manual annotations.In order to simplify and facilitate the process of data annotation,a semi-automatic annotation method based on segmentation refinement of bounding box weak annotation is proposed.The proposed segmentation model is evaluated on two self-built road garbage datasets,which include 7 garbage categories and contain over 3,000 samples.Compared with the original ERFNet model,improved ERFNet has stronger ability of multi-scale feature extraction and feature fusion,with the MIo U rate reaching 88.86% and 68.03% respectively on the two datasets,and shows strong robustness to the image corruptions of the test set.In addition,the inference time per image is 0.12 second for real road images captured by industrial camera of 2 million pixels,which can meet the real-time requirement of practical application of road sweepers.