Abstract Accurate and rapid recognition of dairy cow’s motion behaviours is the key to intelligent perception of its health state. To achieve the recognition of cows’ lying, standing, walking, drinking and feeding behaviours, first, based on the advantage of the efficient feature extraction of EfficientNet, the spatial feature extraction of cow’s video frames was realized. Then, to fully extract the characteristics of different behaviour information of dairy cows, the BiFPN (bidirectional feature pyramid network) was used to realize the efficient fusion of characteristics in the 3–5 layers of EfficientNet. Finally, the behaviour information was sent to the BiLSTM (bidirectional long short-term memory) module, which integrates the attention mechanism to realize the aggregation of video frames in a time series, thus realizing fast and accurate recognition of dairy cow’s motion behaviours. 1009 videos containing 2,270,250 frames of dairy cows in different scenes and postures were collected and tested to compare the performance of the proposed algorithm, four state-of-the-art behaviour recognition algorithms: C3D, VGG16-LSTM, ResNet50-LSTM, and DensNet169-LSTM were carried out. Meanwhile, the precision, recall, accuracy, recall and the average number of frames recognized per second (ANFR) were used to evaluate the performance of the algorithm. Experimental results showed that the behaviour recognition accuracy of the algorithm was 97.87%, which was 4.25% higher than that of the classical ResNet50-LSTM, and the ANFR was 134f/s. In addition, the study was combined with a sliding window to realize the behaviour recognition of undivided single-target dairy cow videos, and the final recognition accuracy could reach 95.20%, showing that the proposed algorithm was effective and could be used for the health status perception and disease prevention of dairy cows.

中文翻译：

---

使用 EfficientNet-LSTM 识别复杂环境中单头奶牛的运动行为

摘要 准确、快速地识别奶牛的运动行为是智能感知奶牛健康状态的关键。为实现奶牛躺、站、走、饮水和喂食行为的识别，首先基于EfficientNet高效特征提取的优势，实现了奶牛视频帧的空间特征提取。然后，为了充分提取奶牛不同行为信息的特征，采用BiFPN（双向特征金字塔网络）在EfficientNet的3-5层实现特征的高效融合。最后将行为信息发送到BiLSTM（双向长短期记忆）模块，该模块集成了注意力机制，实现了时间序列视频帧的聚合，从而实现对奶牛运动行为的快速准确识别。收集并测试了包含 2,270,250 帧奶牛在不同场景和姿势的 1009 个视频，以比较所提出算法的性能，四种最先进的行为识别算法：C3D、VGG16-LSTM、ResNet50-LSTM 和 DensNet169 -LSTM 进行了。同时，通过精度、召回率、准确率、召回率和平均每秒识别帧数（ANFR）来评估算法的性能。实验结果表明，该算法的行为识别准确率为97.87%，比经典的ResNet50-LSTM提高了4.25%，ANFR为134f/s。此外，