基於加速度計之雞隻動作型態識別與活動力估測系統之研究

Abstract

本研究的主要目的為建立一套雞隻動作型態識別與活動力估測之系統。根據裝置在雞隻上的加速度計訊號，來判斷其動作型態，包括走路、啄食、起身與坐下、休息和其他等動作；活動力則以加速度訊號的能量來評估。研究架構分成資料擷取、處理與分析，動作型態識別與活動力估測。資料擷取部分主要整合微機電式三軸加速度計、ZigBee元件與電腦成一無線加速度紀錄器，將此三軸加速度記錄器背負在雞隻的背部，以量測活動的加速度訊號，透過ZigBee將加速度訊號傳回電腦端儲存，同時以DV拍攝活動紀錄；在資料處理與分析方面採用內插法，並將加速度訊號對比相對應的影像資訊，找出不同動作之特徵，本研究採用之特徵，包括三維訊號兩兩間之相關係數、中位數、四分位數間距、峰值數目、頻譜能量、頻譜熵，小波不同頻段之峰值、能量、主頻率等。動作識別部分以動作特徵建立並測試了63種分類模型，以建模資料進行十疊交叉驗證法(10-fold cross-validation)，貝式(Bayesian)網路分類器的辨識準確率為86.1%；以測試資料測驗模型，同種(Homogeneous)資料測試結果，貝式網路的辨識準確率為74.42%；異種(Heterogeneous)資料測試結果，貝式網路的辨識準確率為72.10%，顯示貝式網路之辨識結果較為強健，適合應用在雞隻的動作型態識別。活動力估測係從雞隻活動加速度訊號求得活動資訊，使用小波轉換、活動框架偵測與訊號能量等方法，得到活動框架的平均功率，用以評估雞隻的活動情形與健康程度，達到疾病預警的目的。

The purpose of this paper is to design a system to recognize activity pattern and estimate vitality of chicken. The activity to be recognized includes walk, peck, stand-up and sit-down, rest, and other, etc. The vitality is estimated by energy of acceleration signal. The framework of this study consists of data collection, processing and analysis, activity pattern recognition, and vitality estimation. On data collection, a Wireless Acceleration Logger is designed by integrating MEMS 3-axis accelerometer with ZigBee devices. The three-axis acceleration signals is collected by attaching the Acceleration Logger on the back of chicken, and the collected acceleration signals will be sent to computer by ZigBee. At the same time, a digital video camera is used to record the behavior of chickens. For data processing and analysis, interpolation and wavelet method are used for signal processing. By comparing acceleration signals with the corresponding video clips, the features of various activities could be determined and acquired for further analysis. The features used in this study are correlation coefficients between signals in different axes, median, interquartile range, peak, spectrum energy, spectrum entropy, principal frequency of wavelet bands, amplitude of principal frequency of wavelet bands, and energy of wavelet bands. As of activity recognition, 63 models have been constructed and validated. The accuracy of Bayesian network is 86.10% by 10-fold cross-validation. However, at testing stage, the accuracy of Bayesian network on testing homogeneous dataset is up to 74.42%; the accuracy of Bayesian network with heterogeneous dataset is around 72.10%. The result shows that Bayesian network has the best prediction capability for chicken activity recognition than other models and is also more robust. On vitality estimation, vitality index is estimated from acceleration signals of chicken through wavelet transform, activity frame detection, and the average power of activity frames. The health condition of chicken could be evaluated to achieve the purpose of sickness early warning based on the vitality information.