Comparative Analysis of Human Detection using Depth Data and RGB Data with Kalman Filter: A Study on Haar and LBP Methods

Accurate human detection in video streams with occlusions, illumination variances, and varying distances is crucial for various applications, including surveillance, human-computer interaction, and robotics. This study investigates the performance of two widely used object detection features, Haar-like and Local Binary Pattern (LBP), for detecting human upper bodies in color and depth images. The algorithms are combined with Adaptive Boosting Cascade classifiers to leverage the discriminative power of Haar-like features and LBP texture features. Extensive experiments were conducted on a dataset comprising color images and depth data captured from a Kinect camera to evaluate the algorithms' performance in terms of precision, recall, accuracy, F1-score, and computational efficiency measured in frames per second (fps). The results indicate that when tested on color images, the Haar-Cascade method outperforms LBP-Cascade, achieving higher precision (27.4% vs. 7.8%), recall (49.2% vs. 7.8%), accuracy (21.4% vs. 4.1%), and F1-score (35.2% vs. 7.8%), while maintaining a comparable computational speed (19.07 fps vs. 19.26 fps). However, when applied to depth data, the Haar-Cascade method, coupled with Kalman filtering, demonstrates significantly improved performance, achieving precision (79.3%), recall (79.3%), accuracy (65.8%), and F1-score (79.3%) above 70%, with a computational time of approximately 19.07 fps. The integration of Kalman filtering enhances the robustness and tracking capabilities of the system, making it a promising approach for real-world applications in human detection and monitoring. The findings suggest that depth information provides valuable cues for accurate human detection, enabling the Haar-Cascade algorithm to overcome challenges faced in color image analysis.