Automatic analysis of segmentwise locomotion details of Drosophila larva

Abstract As a simple but beautiful model, the Drosophila larva has been extensively studied to investigate the sensorimotor mechanism by analyzing quantitatively measured behaviors. The current behavior descriptive methods are usually based on the movements of three oversimplified landmarks of the larval body, including the head, tail, and centroid, failing to reveal rich information of multisegment motion. Due to the highly elastic body shapes and sophisticated locomotion, it is challenging to extract body segment movements of a larva to develop a segment-based behavior model. Here, we propose a landmark regression method to automatically detect all the segment joint points of the larva, which can save massive manual efforts. A cascaded regression process is used to precisely estimate segment joint points in each video frame and obtain the detailed segment movement parameters. To train our model, we propose a novel larval body segment dataset, which consists of 3772 images, each image is annotated with 22 segment points. Our method could achieve 99.51% accuracy on the test dataset, which is accurate enough to process typical movement patterns in behavioral experiments for neural coding studies.