Enhanced CAMSHIFT with Perceptual Grouping, Weighted Histogram, Selective Adaptation and Kalman Filtering: Improving Stability and Accuracy in Face Detection and Tracking

Abstract

This paper reports an enhanced CAMSHIFT model for theoretical face detection and tracking. Detection is performed based on a statistical characterisation of skin-colour. The model integrates advanced techniques including Perceptual Grouping, Weighted histogram distribution, Selective Adaptation and with Kalman Filtering to increase the face detection and tracking strategies. Results reveal increased performance in scenarios like hand occlusions, varying illumination, disturbance from multiple faces. The normalized log-likelihood index serves as consistent indicator for face tracking analysis. This new model with Kalman filtering can predict the face’s centroid position and increased tracking stability. This new model has achieved low Mean Absolute Percentage Error (MAPE) both predicted (X^) and (Y^) at 9.32 % and 9.70 % respectively. Low RMSE values of X and Y coordinates reported as 8.3 pixel and 8.8 pixel suggest that the Kalman Filtering predicted values are reliable and accurate. It strongly indicates that on average, the forecast/predicted by Kalman Filtering algorithm deviates from the actual values by low margin and this model is effective in predicting and tracking the face target. Further observation suggests that the x-errors and y-errors have both positive and negative values, suggesting no systematic bias in over- or under-prediction in this developed Kalman Filtering model. This model is a significant advancement in face detection methods, promising improved adaptability and tracking.