基于骨骼关键点的大鲵行为识别模型

Skeletal Keypoint-based Behavioral Recognition Model for Chinese Giant Salamanders

摘要:

针对大鲵夜间活动、生境偏好所导致野外行为数据难以实时精准获取的问题，提出一种基于改进 YOLO 11n-Pose 的关键点检测模型与LSTM-BP级联网络相结合的大鲵行为识别方法。首先构建包含16个关键点的大鲵骨架模型，为提升模型复杂环境下检测精度，提出多维度协同注意力模块（Multi-dimensional collaborative attention，MDCA），通过方向感知、通道增强和空间聚焦 3 个并行分支协同增强特征表示，同时引入Wise-IoU损失函数优化边界框回归质量，以及采用基于L1范数的通道剪枝技术压缩模型体积。在此基础上，提取关键点坐标、关节角度、相对距离和关键点数量等42维特征，形成行为识别的输入数据集，采用LSTM-BP神经网络实现大鲵行为分类，试验结果表明，改进的YOLO 11n-LWM-Pose模型召回率达93.69%、精确率达96.38%、mAP50达95.78%、帧率达88.90 f/s，模型内存占用量仅为6.35 MB。LSTM-BP神经网络识别精确率达96.3%，计算时间仅为3.0 ms。该方法满足野外环境下大鲵行为的实时监测需求，为野生大鲵保护与行为学研究提供有效的技术支撑。

Abstract:

Aiming to address the challenge of obtaining real-time, accurate behavioral data in the wild due to the Chinese giant salamander's nocturnal activity and habitat preferences, a behavioral recognition method that combined an improved YOLO 11n-Pose keypoint detection model with an LSTM-BP neural network was proposed. Firstly, a Chinese giant salamander skeleton model comprising 16 keypoints was constructed. To enhance the model's detection accuracy in complex environments, a multi-dimensional collaborative attention (MDCA) module was proposed, which collaboratively enhanced feature representations through three parallel branches: direction-aware, channel-enhanced, and spatially focused. Concurrently, a Wise-IoU loss function was introduced to optimize bounding box regression quality, and a channel pruning technique based on the L1 norm was adopted to reduce the model's size. Based on this, 42-dimensional features, including keypoint coordinates, joint angles, relative distances, and the number of keypoints were extracted to form the input dataset for behavior recognition. An LSTM-BP neural network was employed to classify giant salamander behaviors. Experimental results showed that the improved YOLO 11n-LWM-Pose model achieved a recall of 93.69%, a precision of 96.38%, an mAP50 of 95.78%, a frame rate of 88.90 f/s, and a memory footprint of only 6.35 MB. The LSTM-BP neural network achieved a recognition accuracy of 96.3% with a computation time of only 3.0 ms. This method met the requirements for real-time monitoring of Chinese giant salamander behavior in the wild, providing effective technical support for the conservation and behavioral research of wild Chinese giant salamanders.