Real-time embedded eye detection system

• Published in: Expert systems with applications 2022-05, Vol.194, p.116505, Article 116505
• Main Authors: Ruiz-Beltrán, Camilo A., Romero-Garcés, Adrián, González, Martín, Pedraza, Antonio Sánchez, Rodríguez-Fernández, Juan A., Bandera, Antonio
• Format: Article
• Language: English
• Subjects: Advanced driver assistance systems; All programmable System-on-Chip; Biometric recognition systems; Digital imaging; Driver fatigue; Embedded systems; Eye detection; Frames per second; Image processing; Object recognition; Power consumption; Real time; System on chip; Viola–Jones algorithm
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2022.116505

The detection of a person’s eyes is a basic task in applications as important as iris recognition in biometric identification or fatigue detection in driving assistance systems. Current commercial and research systems use software frameworks that require a dedicated computer, whose power consumption, size and price are significantly large. This paper presents a hardware-based embedded solution for eye detection in real-time. From an algorithmic point-of-view, the popular Viola–Jones approach has been redesigned to enable highly parallel, single-pass image-processing implementation. Synthesized and implemented in an All-Programmable System-on-Chip (AP SoC), this proposal allows us to process more than 88 frames per second (fps), taking the classifier less than 2 ms per image. Experimental validation has been successfully addressed in an iris recognition system that works with walking subjects. In this case, the prototype module includes a CMOS digital imaging sensor providing 16 Mpixels images, and it outputs a stream of detected eyes as 640 × 480 images. Experiments for determining the accuracy of the proposed system in terms of eye detection are performed in the CASIA-Iris-distance V4 database. Significantly, they show that the accuracy in terms of eye detection is 100%. •A single-pass image-processing redesign of the Viola–Jones approach is proposed.•The classifier core can process more than 750 fps.•An accuracy rate of 100% is obtained using the CASIA-Iris-distance V4 database.•The proposal is integrated into a framework for iris identification on motion.