Temporal action localisation in video data containing rabbit behavioural patterns.

In this paper we present the results of a research on artificial intelligence based approaches to temporal action localisation in video recordings of rabbit behavioural patterns. When using the artificial intelligence, special attention should be paid to quality and quantity of data collected for the research. Conducting the experiments in science may take long time and involve expensive preparatory work. Artificial intelligence based approaches can be applied to different kinds of actors in the video including animals, humans, intelligent agents, etc. The peculiarities of using these approaches in specific research conditions can be of particular importance for project cost reduction. In this paper we analyze the peculiarities of using the frame-by-frame classification based approach to temporal localisation of rabbit actions in video data and propose a metric for evaluating its consistency. The analysis of existing approaches described in the literature indicates that the aforementioned approach has high accuracy (up to 99%) and F1 score of temporal action localisation (up to 0.97) thus fulfilling conditions for substantial reduction or total exclusion of manual data labeling from the process of studying actor behaviour patterns in video data collected in experimental setting. We conducted further investigation in order to determine the optimal number of manually labeled frames required to achieve 99% accuracy of automatic labeling and studied the dependence of labeling accuracy on the number of actors presented in the training data.