小鼠在生理機能上與人類有許多相似處，許多研究皆以其行為作為評估人類疾病的前瞻依據。其中，居家籠監測（Home-cage Monitoring）為主要的觀測項目。然而，長時間觀測小鼠行為需耗費大量人力與時間進行標註。為了降低人力成本並加速行為推估，本研究導入 YOLOv5 （You Only Look Once version 5）深度學習架構，用於自動化辨識小鼠行為並追蹤其空間位置。 本研究設計兩大實驗：首先，針對低光源與曝光的近紅外光環境（Near-Infrared， NIR），建立基於限制對比度自適應直方圖均衡化（Contrast-Limited Adaptive Histogram Equalization， CLAHE）的預處理管線以提升資料品質，並透過遺失率（Missing Rate）與整體準確度評估模型效能。其次，利用 YOLOv5 預測之小鼠位移量進行 3 秒時間段的運動/非運動狀態分類，隨後擬合單通道腦電圖（EEG）自動分期訊號。 研究結果顯示：資料品質擴增表現：於 CLAHE 預處理影像下訓練的模型，其追蹤定位遺失率控制在 0.030，多類別行為偵測之整體準確度達到 0.907。影片運動分類表現，經留一法交叉驗證（Leave-One-Out Cross-Validation， LOO-CV）彙整之數據指出，系統於 16 段日間睡眠影片中的分類準確度為 0.971；相對地，在 16 段清醒影片中的分類準確度則為 0.798。單通道腦波模型表現如下：採用 Convolutional Neural Network和Long Short-Term Memory (CNN + LSTM) 架構的模型在加入 Z-score normalization 處理後，準確率 71.49%；而採用純 CNN 架構的模型在直接使用原始資料推論時，其整體準確率達 83.75%。 這顯示過短時間視窗可能限制了LSTM的時間序列建模能力，而純CNN架構則足以有效捕捉單一時間段的核心頻譜特徵。 總結而言，本研究結合 YOLOv5 運動特徵與單通道電生理指標的多模態睡眠評估框架。該架構在不依賴肌電圖（EMG）訊號的情況下，能重建生理睡眠結構，降低傳統多通道生理量測的侵入性與硬體複雜度，為長時段小鼠研究提供自動化解決方案。;Mice share substantial physiological similarities with humans and are widely used as animal models for investigating disease mechanisms and behavioral phenotypes. Home-cage monitoring enables the observation of natural behaviors under minimally disturbed conditions. However， long-term behavioral assessment requires extensive manual annotation， resulting in considerable labor and time demands. To address these limitations， this study employed the You Only Look Once version 5 (YOLOv5) deep learning framework for automated behavioral recognition and spatial tracking of mice. Two experimental pipelines were developed. First， a Contrast-Limited Adaptive Histogram Equalization (CLAHE) based preprocessing pipeline enhanced near-infrared (NIR) video quality under low illumination and overexposure， evaluated via Missing Rate and overall accuracy. Second， YOLOv5-extracted trajectories provided 3-second epoch displacements for Motion/Non-motion Classification. These behavioral predictions were matched with single-channel electroencephalogram (EEG) sleep identification to evaluate physiological consistency. Results showed that the CLAHE-preprocessed pipeline achieved a 0.030 Missing Rate and a 0.907 behavioral classification accuracy. For visual Motion/Non-motion Classification， Leave-One-Out Cross-Validation (LOO-CV) yielded accuracies of 0.971 across 16 sleep recordings and 0.798 across 16 awake recordings. Concurrently， in single-channel EEG validation， the Convolutional Neural Network + Long Short-Term Memory (CNN + LSTM) model achieved 71.49% accuracy after Z-score normalization， while the CNN model reached 83.75% using raw signals. This discrepancy indicates that the short epoch window restricted the LSTM′s temporal modeling capacity， whereas the pure CNN architecture sufficiently captured the dominant spectral features within individual epochs. In summary， this study established a multimodal Motion/Non-motion framework integrating YOLOv5 behavioral features with single-channel EEG signals. This architecture reconstructs physiological states without electromyography (EMG)， minimizing the invasiveness and hardware complexity of conventional monitoring systems. Ultimately， these findings support the deployment of automated long-term monitoring in rodent biomedical research.
