【作者】 陈颖；

【导师】 苏连成； 刘鑫；

【作者基本信息】 燕山大学 ， 控制工程（专业学位）， 2025， 硕士

【摘要】 小鼠作为神经科学、药理学和行为学研究的重要动物模型,其行为的监测与分析对探究脑神经功能、理解疾病机制及药物研发具有重大意义。然而,传统人工标注效率低下,而以深度学习为代表的数据驱动的行为识别方法具有较高的计算成本和数据依赖性。因此,研究一种基于少样本数据,能高效且精准的自动识别小鼠行为的方法具有重要研究价值。本文针对旷场实验下的小鼠行为识别方法展开研究,通过融合动态特征和静态特征增强关键点检测的鲁棒性,有效应对遮挡。同时,采用12个关键点描述小鼠姿态,通过构建26维特征充分捕捉行为时空信息,提升识别性能。具体研究内容如下:首先,为实现小鼠前景的完整提取,本文设计了一种高效的目标检测技术。通过分段线性化处理增大前景与背景灰度差异,采用OTSU阈值分割算法和主轮廓提取方法初步获取前景区域。在此基础上,提出了一种基于Canny算子的前景补全策略,通过填充断裂边缘线与前景区域的内部连接点,修复前景丢失部分。实验效果表明,该方法能稳定、准确地提取小鼠的完整前景区域,有效去除背景干扰。随后,本文开发了一种基于多特征融合的小鼠姿态关键点检测方法,不依赖于标注数据。基于几何、灰度信息、轮廓曲率等静态特征初步提取关键点,包括小鼠鼻尖、两耳中点、左耳中心点、右耳中心点、颈部、左前肢、右前肢、重心、臀部、左后肢、右后肢、尾基点。利用帧差图像并结合光流金字塔方法,捕捉运动信息,以补偿和修正静态关键点位置。实验结果表明,该方法对小鼠身体各部位平均检测准确率达到94%,相较于Deep Lab Cut和SLEAP方法略有提升,具有一定的先进性。最后,构建了一个26维时空特征集,用于减低训练成本并实现小鼠行为的精确分类。通过计算姿态关键点间的距离、角度以及目标检测框面积量化空间特征,同时提取关键点帧间位移、速度及加速度等运动参数来构建时间特征。二者组合以全面描述小鼠姿态变化,并以支持向量机作为分类器,对扶壁直立、悬空直立、修饰、跳跃、行走等经典行为进行识别。实验结果显示,所提方法的平均行为识别准确率达到91.7%,略高于Deep Ethogram方法,体现了其在少样本场景下的高效性和可靠性。更多还原

【Abstract】 As a crucial animal model in neuroscience,pharmacology,and behavioral research,the monitoring and analysis of mouse behavior are of great significance for investigating brain neural functions,understanding disease mechanisms,and facilitating drug development.However,traditional manual annotation is inefficient,while data-driven behavior recognition methods,represented by deep learning,suffer from high computational costs and heavy data dependency.Therefore,developing an automated mouse behavior recognition method that is both efficient and accurate while relying on few-sample data holds substantial research value.This study focuses on the mouse behavior recognition under open field test,enhancing the robustness of keypoint detection by integrating dynamic and static features to effectively address occlusion issues.Additionally,12 keypoints are employed to describe mouse posture,and a 26-dimensional feature set is constructed to comprehensively capture spatiotemporal behavioral information,thereby improving recognition performance.The specific research contributions are outlined as follows:First,to achieve accurate extraction of the mouse foreground,this work designs an efficient object detection technique.By applying piecewise linearization to amplify the gray-scale contrast between the foreground and background,followed by preliminary foreground extraction using the OTSU threshold segmentation algorithm and primary contour extraction.Building upon this,a foreground completion strategy based on the Canny operator is proposed,which repairs missing foreground regions by filling in broken edge lines and internal connection points.Experimental results demonstrate that this method can stably and accurately extract the complete mouse foreground while effectively eliminating background interference.Subsequently,this paper develops a multi-feature fusion-based keypoint detection method for mouse posture estimation,eliminating reliance on annotated data.Initial keypoints—including the nose tip,midpoint between ears,left ear center,right ear center,neck,left forelimb,right forelimb,center of mass,hip,left hindlimb,right hindlimb,and tail base—are extracted based on static features such as geometry,grayscale information,and contour curvature.Motion information is then captured using frame differencing combined with an optical flow pyramid method to compensate for and refine static keypoint positions.The experimental results show that the average detection accuracy of the proposed method for various parts of the mouse body reaches 94%,slightly surpassing the Deep Lab Cut and SLEAP,demonstrating certain superiority.Finally,a 26-dimensional spatiotemporal feature set is constructed to enable precise classification of mouse behaviors while reducing training costs.Spatial features are quantified by calculating distances and angles between posture keypoints,as well as the area of the bounding box,while temporal features are constructed by extracting inter-frame displacements,velocities,and accelerations of keypoints.These features are combined to comprehensively describe mouse posture dynamics,with a support vector machine(SVM)serving as the classifier to recognize classic behaviors such as wall rearing,suspended rearing,grooming,jumping,and walking.Experimental results show that the proposed method attains an average behavior recognition accuracy of 91.7%,slightly exceeding that of the Deep Ethogram method,highlighting its efficiency and reliability in scenarios with limited samples.更多还原