Research On Semantic Segmentation Technology Of Road Scene Based On Deep Learning

The road detection algorithms which are based on the traditional computer vision methods suffer from the problems such as low generalization ability,difficulty in setting proper parameters,and low detection accuracy.The existing deep learning semantic segmentation pre-trained models could not completely take into account the 3 basic task requirements,which are high detection accuracy,low resource occupancy and high frame rate,of our smart garbage sweeper projects as well.This work designs and builds a new Campus Road Dataset which collects images that describe the working conditions of smart garbage sweepers.To satisfy the needs of this project,a semantic segmentation neural network which uses the Mobile Netv2 as the backbone structure is trained on the Cityscapes dataset and the Campus Road Dataset from scratch and the trained model implements the segmentation task of the road.And a structured street lane detection method which is based on traditional image processing algorithm is exploited to locate the working regions of the sweepers.With the help of these two modules,the basic demands of the smart garbage sweeper projects are achieved in terms of detection speed and accuracy.In order to decrease the resource occupancy,our work combines a discrete cosine transform with channel selection module with the backbone structure,successfully reducing the total trainable parameters of the networks.Additionally,this work designs a category-based focal loss which aims to increase the average detection precision of the model.The experiment result shows that our proposed loss function increases the semantic category with the lowest Io U by 6% and the average Io U of all categories by 1%.The generalization ability of category-based focal loss is exhaustively tested on 3 different network backbones and 2 benchmark semantic segmentation datasets and the comparison experiment between our proposed focal loss function,original focal loss function,and cross entropy loss function validates that categorybased focal loss could achieve higher performance in different models and datasets.