Real-Time Automated Hand-Vote Counting System

Abstract

Manual vote counting has come under increasing scrutiny as election experts and critics question its accuracy, efficiency, and susceptibility to human error. This issue becomes more pronounced as the number of votes increases. In response to this problem, researchers have explored alternative approaches, such as automation, to address the challenges associated with the traditional “Show of Hands” method of voting. By employing computer vision techniques, the researchers have conceptualized the development of algorithmic intelligence using the computational power of Raspberry Pi. This system can recognize open-palm hand patterns and offers a reliable and efficient mechanism for expediting the process of school elections. Specifically, this study employs CNN-based object recognition algorithms, namely the YOLOv5 and YOLOv8 nano models. These models were trained for 100-300 epochs using the HAGRID ‘Stop Sign’ datasets and researchers’ self-produced datasets containing multiple images of raised hands. Performance metrics, including Mean Average Precision, Confusion Matrix, Inference Speed, and Model Weight were considered and monitored to evaluate the effectiveness of the models. As a result, the training stage yielded an 82.6% Mean Average Precision with an IoU (Intersection over Union) threshold of 0.50:0.95. This achievement indicates a satisfactory performance in object detection, surpassing the threshold typically considered as indicative of good performance.