Abstract
Gesture interaction is one of the novel human-computer interaction methods for smart TVs. Addressing the issues of false detection and high detection costs in gesture recognition algorithms for gesture interaction, this paper proposes the YOLOv8n-Remote Finger (YOLOv8n-RF) algorithm for dynamic remote control finger detection. This algorithm utilizes the CRVB-DSConvEMA module in the feature extraction network, adopts the SPPF-DSConvEMA module in the downsampling process, and introduces BiFPN in the Neck layer. Experiments conducted on the self-made Remote Finger dataset and the public HaGRID dataset demonstrated that, compared to the YOLOv8n algorithm, the proposed YOLOv8n-RF algorithm achieved an improvement in mean Average Precision (mAP) by 1.23% and 0.84%, respectively. Additionally, the model size was reduced by 2.49 M, the GFLOPs were decreased by 1.7, and the false detection rate was lowered by 22%. The YOLOv8n-RF algorithm meets the requirements of low cost and low complexity, which contributes to reducing false control operations on smart TVs.