Hybrid features for RGB-D object tracking using deep learning
Abstract
Tracking objects in red, green, blue, and depth (RGB-D) video remains challenging because visual appearance can change under occlusion, illumination variation, and motion blur, while depth frames may contain missing or noisy values. This study proposes a hybrid RGB-D object-tracking framework that combines convolutional neural network (CNN) features with local binary patterns (LBP) for texture and histogram of oriented gradients (HOG) for shape, using a Kalman filter and a deep appearance model to estimate object locations across frames. The method is evaluated on the Digital Twin Tracking Dataset (DTTD), comprising 100 videos with annotated ground-truth tracks. The experimental protocol uses an 80:20 training/testing split, 100 training epochs, a batch size of 32, and rotation and horizontal-flip augmentation. On the reported benchmark, the proposed approach achieves a success rate (SR) of 88.9%, a precision rate (PR) of 91.5%, and a root mean square error (RMSE) of 3.9 pixels, at 22 frames per second (FPS). These results indicate that combining learned semantic representations with handcrafted texture and shape descriptors improves tracking performance over the individual feature-based baselines, particularly under occlusion and illumination changes.
Keywords
Convolutional neural networks; Deep learning; Feature fusion; Hybrid features; Kalman filter; Object tracking; RGB-D tracking
Full Text:
PDFDOI: https://doi.org/10.11591/eei.v15i6.12903
Refbacks
- There are currently no refbacks.

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
Bulletin of Electrical Engineering and Informatics (BEEI)
ISSN: 2089-3191
,
e-ISSN: 2302-9285
This journal is published by the
Institute of Advanced Engineering and Science (IAES)
.