Social behavior analysis of Drosophila larvae via motion activity recognition

We aim to analyze the social behavior of Drosophila larvae through automated identification of their motion activity pattern. Social activity recognition is an essential step in identifying the correlation between specific neuronal structures in their brain to the social behavior of the species. Larvae videos are obtained using low-cost overhead video; subsequently, the motion features of the larvae generated by computing a gradient weighted optical flow. These features are used to create a global histogram of gradient-weighted optic flow over a specific time frame. The method for estimating optical flow is based on an implementation by Gautama and Van Hulle that expands on the classical Horn and Schunck method by using spatial Gabor filter pair to find the intensity gradient and recurrent neural networks to find velocity. Activity classification is obtained by using a support vector machine and yields 82% overall accuracy in classification.