Bio-logging technology is becoming an ever more common tool for persistent monitoring of people and animals in their natural environment. As a result, the volume and type of information collected by these embedded sensing systems continues to increase, making algorithms that can accurately and efficiently classify and parameterize behavior from sensor data essential. How best to extract information from multiple sensors remains an open question. The problem becomes more challenging in cases where only sparse concurrent human observations of the behavioral states are available to train the algorithm. In this work, the authors present a systematic method to perform feature generation, feature selection and state classification from representative data collected from an example species — bottlenose dolphins. This approach includes methods for evaluating window size selection during feature generation and the identification of specific feature sets that maximize classification performance. Additionally, the proposed framework incorporates information about state transition probabilities to further improve classification accuracy. Bio-logging sensor and video data for the analysis are collected from free-swimming dolphins at Dolphin Quest Oahu. The concurrent video data is scored by a human expert to create a set of observed behaviors. Results demonstrate that the algorithm is able to classify behavior with a high level of accuracy (> 90 percent) with 16 features and a window size of 0.6 seconds. Robustness of the proposed approach is evaluated by reducing the training data by 80 percent. The resulting classification accuracy is still above 87 percent. These results serve as the foundation for classification algorithms that can be used with data collected from animals where behavioral states can only be observed sporadically.

This content is only available via PDF.
You do not currently have access to this content.