Abstract: In order to realize the fast and accurate positioning and identification of obstacles in the working environment by lawn mowing robot with limited computing resources, an obstacle detection method of mowing robot based on improved YOLOv5s deep learning model with filter pruning is proposed. Firstly, the YOLOv5 model uses a layered residual structure to represent multiscale features with finer granularity, and the network receptive fields are added. In addition, SE module is added to the tail of the residual block to recalibrate the feature map. Secondly, the filter pruning is performed for the improved algorithm. Finally, the relevant data sets were established for common obstacles in the working environment of lawn mowing robot, and the improved YOLOv5s after pruning was used as a deep learning model for detection. Experimental results show that the size of the improved YOLOv5 model is reduced by 18.8%, and the mAP is increased by 0.1%. After pruning the improved YOLOv5 model, the computational amount and the model size are reduced by 36.6%, 33.3%, and 1.9 ms, respectively, compared with the original model. After pruning, the mAP of the final model is 1.3%, 9.5%, 5.8% and 22.1% higher than that of YOLOv4, YOLOV4-tiny, YOLOv3 and YOLOV3-tiny, respectively.

Key words: deep learning, mowing robot, object detection, model pruning

摘要： 为实现割草机器人在计算资源有限的情况下快速、准确地定位并识别工作环境中的障碍物，提出一种基于滤波器剪枝的改进YOLOv5s深度学习模型的割草机器人工作环境下障碍物的检测方法。首先，将YOLOv5模型中的Bottleneck残差块改为分层残差结构，以更细粒度地表示多尺度特征，同时增加网络感受野；另外，在残差块尾部加入SE模块，用来对特征图重新标定；其次，对改进后的算法进行滤波器剪枝；最后，针对割草机器人工作环境中的常见障碍物建立相关数据集，并使用剪枝后改进YOLOv5s作为深度学习模型进行检测。试验结果表明：改进后的YOLOv5模型大小减少188%，mAP增加0.1%。对改进YOLOv5模型进行剪枝后，比原YOLOv5模型计算量降低36.6%，模型大小降低333%，推理速度减少1.9 ms。剪枝后本文模型的mAP值分别比YOLOv4，YOLOv4-tiny，YOLOv3，YOLOv3-tiny高1.3%，9.5%，5.8%，22.1%。

关键词: 深度学习, 割草机器人, 目标检测, 模型剪枝