Abstract
Abstract: Environmental information perception has been one of the most important technologies in agricultural automatic navigation tasks, such as plant fertilization, crop disease detection, automatic harvesting, and cultivation. Among them, the complex environment of a field road is characterized by the fuzzy road edge, uneven road surface, and irregular shape. It is necessary to accurately and rapidly identify the passable areas and obstacles when the agricultural machinery makes path planning and decision control. In this study, a lightweight semantic segmentation model was proposed to recognize the unstructured roads in fields using a channel attention mechanism combined with the multi-scale features fusion. Some environmental objects were also classified into 12 categories, including building, person, vehicles, sky, waters, plants, road, soil, pole, sign, coverings, and background, according to the static and dynamic properties. Furthermore, a mobile architecture named MobileNetV2 was adopted to obtain the image feature information, in order to reduce the model parameters for a higher reasoning speed. Specifically, an inverted residual structure with lightweight depth-wise convolutions was utilized to filter the features in the intermediate expansion layer. In addition, the last two stages of the backbone network were combined with the Hybrid Dilated Convolution (HDC), aiming to increase the receptive fields and maintain the resolution of the feature map. The hybrid dilated convolution with the dilation rate of 1, 2, and 3 was used to effectively expand the receptive fields, thereby alleviating the "gridding problem" caused by the standard dilated convolution. A Channel Attention Block (CAB) was also introduced to change the weight of each stage feature, in order to enhance the class consistency. The channel attention block was used to strengthen both the higher and lower level features of each stage for a better prediction. In addition, some errors of semantic segmentation were partially or completely attributed to the contextual relationship. A pyramid pooling module was empirically adopted to fuse three scale feature maps for the global contextual prior. There was the global context information in the first image level, where the feature vector was produced by a global average pooling. The pooled representation was then generated for different locations, where the rest pyramid levels separated the feature maps into different sub-regions. As such, the output of different levels in the pyramid module contained the feature maps with varied sizes, followed by up sampling and concatenation to form the final output. The results showed that the objects in the complex roads were effectively segmented with Pixel Accuracy (PA) and Mean Pixel Accuracy (MPA) of 94.85% and 90.38%, respectively. Furthermore, the single category pixel accuracy of some objects was more than 90%, such as road, plants, building, waters, sky, and soil, indicating a higher accuracy, strong robustness, and excellent generalization. An evaluation was also made to verify the efficiency and superiority of the model, where the mean intersection over union (MIoU), segmentation speed, and parameter scale were adopted as the indexes. The FCN-8S, SegNet, DeeplabV3+ and BiseNet networks were also developed on the same training and test datasets. It was found that the MIoU of the model was 85.51%, indicating a higher accuracy than others. The parameter quantity of the model was 2.41×106, smaller than FCN-8S, SegNet, DeeplabV3+, and BiseNet. In terms of an image with a resolution of 512×512 pixels, the reasoning speed of the model reached 8.19 frames per second, indicating an excellent balance between speed and accuracy. Consequently, the lightweight semantic segmentation model was achieved to accurately and rapidly segment the multiple road scenes in the field environment. The finding can provide a strong technical reference for the safe and reliable operation of intelligent agricultural machinery on unstructured roads.
DownLoad: