【摘要】 由于山地果园运输机立地条件差，实时作业信息的获取、反馈、集中化管理较为困难，为了解7SYDD–200型山地果园单轨运输机搭载货物情况，合理调度运输装备，建立了基于改进的YOLOv5s模型的运输机搭载柑橘果筐的检测方法：在果园自然光环境下使用RGB相机(HSK–200)采集运输机搭载柑橘果筐的图像数据；建立和优化YOLOv5s模型，部署至嵌入式设备，实现对搭载过程中的“空果筐”“柑橘”“满果筐”状态的检测。在模型的颈部网络引入CBAM注意力机制，加强模型提取语义信息的能力，解决检测过程中出现的“双重标签”的问题，使用批归一化(BN)层稀疏的尺度因子衡量各通道对模型的表征能力，并对表征能力弱的通道进行剪枝压缩，以克服基模型YOLOv5s检测速度慢的问题，通过多尺度训练策略对模型进行微调，提高模型检测准确率。试验结果表明：改进YOLOv5s模型的检测方法在柑橘搭载数据集上平均精度均值(m AP)为93.3%；模型的浮点数运算量和大小分别为9.9GFLOPs和3.5 MB，比YOLOv5s的提高60.3%和21.3%；在嵌入式平台Jetson Nano部署，其检测速度为78 ms/帧。更多还原

【Abstract】 Due to the poor site conditions of mountainous orchard monorail transporter, it is difficult to obtain, feedback and centralized management of real-time operation information. In order to monitor the proceeding of deliveries by 7SYDD-200 mountainous orchard monorail transporter and reasonably dispatch transportation equipment, the detection method of citrus fruit basket carried by the transporter is established based on the improved YOLOv5s model. Images of the citrus fruit baskets carried by the transporter were collected by the RGB camera of HSK-200 under the natural light environment of mountainous orchards. The YOLOv5s model was established and optimized, which was deployed into the embedded device to detect the states of “empty fruit basket”, “citrus” and “full fruit basket” during the loading process.convolutional block attention module(CBAM) is introduced into neck network of the model to strengthen the ability to extract semantic information and solve the problem of “double labels” in the detection process. The sparse scale factor of the batch normalization(BN) layer was used to measure the representation ability of each channel of the model. The channels with weak representation ability were pruned and compressed to overcome the problem of slow detection speed of the model based on YOLOv5s. The multi-scale training strategy is used to fine-tune the model to improve the detection accuracy. The test results show that the mean average precision of the improved detection method is 93.3% on the fruit dataset. The floating point operation and the size of the improved models were 9.9 G and 3.5 M, respectively, which were 60.3% and 21.3% higher than that of YOLOv5s. The detection speed of the improved model was 78 ms/img, when it was deployed into the Jetson Nano embedded platform.更多还原