Automated Livestock Vocalisation Detection in Farm Acoustic Environments

Abstract

<p>Precision Livestock Farming (PLF) is the application of process engineering principles to livestock management, utilising automation, sensor-based monitoring, and intelligent systems to augment the expertise of livestock producers, and to facilitate the continuous monitoring of livestock biological responses. Livestock vocalisations have been shown to contain a wealth of information pertaining to animal welfare and health, presenting a significant opportunity to facilitate the noninvasive observation of livestock. The automated detection of livestock vocalisations has been demonstrated in numerous applications, such as welfare monitoring, disease diagnosis, and reproduction efficiency.</p> <p>Recent advancements in machine learning have accelerated the progress of acoustic identification technologies, and this thesis will explore their application to the problem of livestock vocalisation detection. An automated livestock detection algorithm is proposed, utilising novel features derived from a Discrete Wavelet Transform (DWT), combined with a Support Vector Machine (SVM) approach to machine learning. A computationally efficient energy-based thresholding method was employed to facilitate automated segmentation. To date, a complete end-to-end realtime vocalisation detection system has not been demonstrated in the literature. This thesis culminates in the creation of prototype real-time automated vocalisation detection hardware, deployed using a low-resource single board computer (SBC).</p> <p>To investigate the plausibility of an automated livestock detection system, MelFrequency Cepstral Coefficients (MFCCs) were used as input features for an SVM model to accurately classify sheep vocalisations. To improve on the preliminary results obtained, compact DWT-based acoustic features were proposed. The developed features were compared to MFCCs, both in terms of classification performance and computational timing. It was found that DWT-based features provided similar discriminatory ability to MFCCs, but a marked difference in computational timing was observed. DWT-based features were consistently faster to compute, and less variable in computational time, make them an ideal candidate for real-time vocalisation detection applications. To test the algorithm’s ability to be rapidly applied to different livestock-related vocalisations, 3 datasets were used for performance evaluation, targeting sheep, cattle, and Mareema Sheepdog vocalisations. The high classification results obtained indicated that the algorithm is multi-purpose in nature and can be successfully retrained to detect different vocalisation types.</p> <p>With the successful development of an accurate classification component, functionality was extended into automated segmentation using an energy-based thresholding approach to acoustic event identification. A sophisticated codebase was developed in the C-programming language, capable of operating in an offline capacity. 700 hours of field recorded audio data underwent automated segmentation to identify the periods of highest acoustic activity. A test dataset containing sound events temporally dissimilar to the training data was created, allowing for the evaluation of the algorithm’s ability to operate in a changing acoustic environment. Results obtained from the test dataset revealed a highly accurate vocalisation detection algorithm, capable of successfully extracting and identifying sheep vocalisations. Although the algorithm performed well at classifying negative acoustic events, it was found to struggle with sounds possessing a similar spectral content and envelope to the target class (e.g., crow vocalisations). Following the successful development of the automated detection algorithm, all field recorded data was processed to classify segmented sound events, revealing the ratio of sheep vocalisations contained within each period.</p> <p>The codebase developed for offline operation was extended to real-time deployment using the Advanced Linux Sound Architecture (ALSA). Prototype real-time livestock vocalisation detection hardware was produced and tested using pre-recorded audio data playback at varying distances from the capture sensor. During experimentation, the real-time system demonstrated high performance in detecting sheep vocalisations, and correctly rejected the majority of non-vocalisation acoustic events. The system experienced no data losses and did not enter any error states, indicating that the developed algorithm is computationally efficient enough to run on lowresource SBC hardware.</p>