Intelligent sweeper vehicle is gradually applied to human life, in which the accuracy of garbage identification and classification can improve cleaning efficiency and save labor cost. Although Deep Learning has made significant progress in computer vision and the application of semantic network segmentation can improve waste identification rate and classification accuracy. Due to the loss of some spatial information during the convolution process, coupled with the lack of specific datasets for garbage identification, the training of the network and the improvement of recognition and classification accuracy are affected. Based on the Unet algorithm, in this paper we adjust the number of input and output channels in the convolutional layer to improve the speed during the feature extraction part. In addition, manually generated datasets are used to greatly improve the robustness of the model. Next, we select the Softmax as the activation function to adjust the classification probability of each category. The cross-entropy function, selected as the loss function, is used to evaluate the fitting ability of the model. We adjust the stride and the size of the feature map in the convolution process to reduce the amount of calculation. The k-fold cross-validation makes full use of the dataset to better optimize the model. Compared with Unet algorithm, the data shows that the improved algorithm can extract more effective features for classification prediction, and simplify the network for parameter selection and optimization, and improve the training speed by 10%. The recognition accuracy is increased by 18% while ensuring the speed, and it has higher MPA and MIoU. This algorithm has good generalization ability to deal with different test sets, which improves the efficiency of the intelligent sweeper vehicle and provides a reference for the design of the semantic segmentation model of garbage classification at the same time.