Navigation path recognition for tea plantation mobile harvesting robots based on improved Deeplabv3plus

In complex tea plantation environments, efficient navigation path recognition algorithms are essential to guarantee the autonomous operation of mobile harvesting robots. However, the relevant path recognition algorithms are poor in accuracy and real-time, making it difficult to identify complex tea plantation paths quickly and accurately, which affects the efficiency and robustness for autonomous operation of tea plantation mobile harvesting robots. To address the above problems, firstly, this paper selected Deeplabv3plus as the model framework for path recognition, and employed lightweight Mobilenetv2 as the backbone network of the model to optimize the number of arithmetic parameters and the speed of the model. Secondly, an efficient CE_ASPP feature extraction module was designed to construct a multi-scale depth-separable convolutional structure to expand the feature receptive field of the model. In the decoding layer, this paper designs the CMG feature extraction module, which integrates the shallow features of the first three bottleneck layers of the backbone network, is fused to enhance the semantic information of the features; and the feature extraction capability was improved by the CBAM attention mechanism and the ECA attention mechanism. Finally, the scanning method was employed to extract the central pixel, while the RANSAC algorithm facilitates the fitting of the navigation path. The experimental results showed that the Accuracy, mPA and MIoU of the improved Deeplabv3plus model were 95.7 %, 97.64 % and 93.51 %, respectively. Compared to the original model, the number of operational parameters and memory footprint of the improved model were 4.35×10⁵ and 16.60MB respectively, which significantly reduces the computational costs. The average inference time and the average detection frame rate were 48.99ms and 20.44FPS respectively. The results showed that the improved Deeplabv3plus has better detection accuracy and robustness for complex tea plantation paths, providing important solutions for autonomous operations of tea plantation mobile harvesting robots.