Automatic Horse Blink Detection Using Computer Vision and Deep Nets

Measurements of dopaminergic activity in the central nervous system provide valuable information about animal health and welfare. In horses, it has been shown that blink rate is correlated to dopaminergic activity and can be used as a non-invasive biomarker. In this paper, we propose two new algorithms for video-based automatic blink detection in horses. The first algorithm employs an OpenCV object tracker to localize the eye, and detects blinks from local color changes over successive frames. The second algorithm is based on a neural net classifier which categorizes each video frame into either “eye is open” or “eye is closed” categories. It then clusters “eye is closed” frames into distinct blink events. Both algorithms run a post-processing method to improve prediction accuracy by removing outliers and merging neighbouring clusters that belong to the same blink event. The test data set consisted of eight RGB video recordings from three healthy horses moving freely in outdoor environment. Our results show that the first algorithm had better accuracy (81% > 31%, p<0.01) and lower error rate (27% < 69%, p<0.01) than the second algorithm. This study is part of an ongoing work to develop a cost-effective, non-invasive and automated health monitoring system for horses and other bovine animals.