Real-Time Isolated Sign Language Recognition

Various researchers have proposed models with high recognition rates for sign language recognition. Moreover, with the improvement of GPU processing power, more and more proposals for real-time processing have been made. We have proposed methods to improve the recognition rate by adding a new approach to the Skeleton Aware Multi-modal Sign Language Recognition (SAM-SLR) using the Ankara University Turkish Sign Language (AUTSL) dataset. This SAM-SLR uses four modalities and fuses their results to achieve a high recognition rate. However, when processing in real-time, the four modalities require more processing and extended response time than other methods. We propose a method for real-time processing while maintaining the original evaluation value based on the published code of this SAM-SLR. Our proposed method is faster by reducing processing, parallel processing, utilizing internal memory, etc. As a result, the average response time was 0.7248 s for the proposed processing method and 4.4013 s for the non-accelerated serial processing method, an 83.5% improvement with an average processing speed of 6.07 times.