Visual navigation path recognition method for tea plantation based on improved DeeplabV3+

Abstract

• The improved DeeplabV3+ for tea plantations to recognize the navigation path. • An efficient MSE_ASPP feature extraction module for expanding the feature receptive field. • A SAM feature extraction model for enhancing the model's feature extraction capability. • A scanning method and a Huber regression model for fitting the navigation path based on the mask image of tea bushes. The accurate recognition of navigation paths is imperative for the autonomous navigation of mobile robots in tea plantations. However, traditional path extraction algorithms have been shown to exhibit poor detection accuracy and generalization capabilities. These issues result in difficulties in accurately extracting tea bush paths, thereby compromising the navigation robustness of robots. To address the aforementioned issues, this paper proposes a tea plantation visual navigation path recognition method based an improved DeeplabV3+. The MobileNetv2 is utilized as the backbone network to optimize computational cost. Multi-scale deep separable convolution structures and ECA feature extraction branch have been introduced to construct an efficient MSE_ASPP feature extraction module, thereby expanding the feature receptive field. In the decoding layer, The SAM feature extraction module is designed to enhance semantic information by fusing the shallow features from the first three layers of the backbone network. Furthermore, this paper employs a scanning method and a Huber regression model in combination with segmented mask regions to fit the navigation paths. The experimental results show that the MPA, Accuracy, and MIOU of the improved model reached 94.96%, 94.27%, and 90.03%, respectively. In comparison with the DeeplabV3+, the computational parameters and memory footprint of the improved model reached 4.14 × 10 6 and 15.81 MB, respectively, greatly reducing the computational cost of the original model and the average inference time reached 44.86 ms. Furthermore, the mean heading angle error of the robot reached 4.10°. The experimental results demonstrate that the improved model exhibits excellent robustness and is capable of extracting precise tea bush navigation path, thus providing a significant solution for visual navigation tasks of mobile robots in tea plantations.