Effectiveness of gyroscopes and triaxial accelerometers paired with deep learning algorithms in detecting dairy camel behavior

Abstract

Nowadays, the behavior of dairy camels in intensive systems has received little attention. This study is the first to use wearable sensors to predict camel behavior, filling a knowledge gap by providing information into their activities. A novel system using triaxial accelerometer and gyroscope sensor was developed to monitor and predict the behavior of six female Maghrebi dairy camels. Using a 10-second time window for data collection, this research effectively distinguishes between key behaviors such as feeding, ruminating, resting, and walking. Various deep learning techniques, including convolutional neural networks combined with long short-term memory (ConvLSTM), dense layer convolutional neural networks, and standalone long short-term memory (LSTM) networks, were used to analyze the data. Results indicate that, under our conditions, camels spent most of the study time feeding, with 33134 occurrences (444 minutes), followed by ruminating 11611 times (156 minutes) and resting 10487 times (140 minutes). In contrast, walking and drinking were much less frequent, with 2627 and 368 occurrences, respectively. The dense-layer CNNs achieved the highest predictive performance with an overall accuracy of 84 %. This model predicted feeding with 89 % accuracy, resting with 67 %, ruminating with 92 %, and walking with 12 %. Following closely, the ConvLSTM model attained an accuracy of 83 %, predicting feeding at 85 %, resting at 76 %, ruminating at 87 %, and walking at 18 %. The LSTM model had a slightly lower overall accuracy of 78 %, predicting feeding at 81 %, resting at 66 %, ruminating at 87 %, and walking at 8 %. In the ConvLSTM model, resting was frequently confused with feeding and ruminating, while walking was often misclassified as feeding. Similarly, the Convolutional with Dense Layers model misclassified resting and walking as feeding, and ruminating as resting. The LSTM model showed similar issues, with resting and walking misclassified as feeding, and ruminating often confused with both feeding and resting. This study highlights the potential of accelerometer and gyroscope sensors as effective tools for assessing camel behavior in intensive systems. The dense layer CNN model showed the best predictive performance, with feeding and rumination behaviors being the most accurately classified. However, walking remained difficult to predict across all models. This is probably due to the limited locomotion of camels in intensive dairy systems. These findings provide a basis for improving automated behavioral monitoring in dairy camels, supporting improved welfare and optimized management in intensive farming systems.