Real-Time behaviour classification techniques in low-power animal borne sensor applications

[摘要] ENGLISH ABSTRACT: The ability to study animal behaviour is important in many areas of science, including behaviouralecology, conservation and precision farming. These studies typically employ biotelemetrytags attached to animals that collect raw sensor data from tri-axial accelerometers.However, conventional animal behaviour classification techniques are performed offline as apost-processing step and does not provide real-time data analysis. Furthermore, the lifespanof such tags is constrained by their power and memory usage, which are often limiting factorswhen performing behavioural studies for extended periods of time. The focus of this projectwas to investigate methods to possibly mitigate these limitations. The main contributionsof the work set out in this dissertation are three-fold. First, a novel embedded automaticbehaviour classification system which captures and automatically classifies three-dimensionalaccelerometer data in real-time is presented. All computation occur on specially designedbiotelemetry tags while attached to the animal. This allows the probable real-time behaviourto be transmitted continuously, thereby providing an enhanced level of detail and immediacy.As a result of the sustained and serious rhinoceros poaching in South Africa, the behaviourclassification system was developed to assist with activities combating this problem. An onboardlinear support vector machine with 11 features achieves an accuracy of 99:61% amongthree behavioural classes (standing, walking and lying down). Stock theft is another significantproblem as experienced in the agricultural sector. The behaviour classification systemwas, therefore, also implemented for sheep. In this case, logistic regression with 34 featuresachieves a classification accuracy of 89:59%among five behavioural classes (standing, walking,grazing, running and lying down). The estimated behaviour was established approximatelyevery 6:5 s and transmitted to a receiver station for both rhinoceros and sheep. Secondly, anovel energy-aware feature and model selection technique is presented. A greedy sequentialfeature selection algorithm was utilised to minimise a cost function that weighs the energy expense of adding specific features with the change in classification error afforded by the features.In addition, the energy expense of specific classification techniques are considered inselecting the optimal models, which is often neglected in literature. Our technique, therefore,favours both classifiers and features which are less energy expensive to compute duringruntime. It is shown that, for the rhinoceros dataset, a random forest classifier with two features is selected as optimal, achieving an overall classification accuracy of 99:33%. Extractingthe features and performing classification consumes 363 times less energy, while only sacrificing0:28% in accuracy when compared to the 99:61% achieved with the unconstrainedsystem. For the sheep dataset, a linear support vector machine with nine features achievesan 88:40% classification accuracy. Extracting the features and performing classification consumes6.8 times less energy, at a cost of 1:19% in accuracy compared to the 89:59% achievedwith the unconstrained system. Finally, the reduced power requirements and memory usagebenefits of the embedded behaviour classification system were considered. Experiments usingthe biotelemetry tags demonstrated a 14-fold reduction in energy consumption and a 234-foldreduction in memory usage when classification was performed on the tag vs. processing rawdata subsequent to transmission. It is concluded that real-time behavioural updates can beachieved by means of embedded behaviour classification with the technique significantly reducingthe total energy consumption and memory requirements of the device. This enableslong-term behavioural studies in applications such as the conservation of rhinoceros, whichis a critically endangered species. It is also very applicable to precision farming applications.Moreover, this technique can be applied to general embedded machine learning applicationsemployed in smart phones, smart watches and sensors within the internet of things.

[发布机构] Stellenbosch University