Blink detection is an important task for human-computer interaction and behavior analysis. Although there is previous research regarding drowsiness detection, computer vision syndrome, and computer access by disabled patients, these have certain limitations for their algorithm’s accuracy due to a wide range of acquisition. Particularly, head movements, scene conditions, and the number of people in a frame present the main limiting factors. This paper proposes a low latency algorithm based on image processing and a convolutional neural network (CNN). The first technique is used to simplify the amount of computational cost by reducing the input data of the CNN. Then, the CNN is used to classify whether a specific frame is in an ‘open’ or ‘closed’ eye state. As this proposal was tested in a development board, limited CPU specifications and a reduced image database were considered for the CNN architecture and its training. The algorithm was tested using a CSI camera and a Jetson Nano 4 GB development board, obtaining a 99.5% accuracy for blink detection.