A Comparative Analysis of Traditional and CNN-Based Object Recognition Techniques in Robotics

This paper presents a comparative study of traditional object recognition methods and CNN-based algorithms, applied within the scope of robotics. Traditional techniques such as SIFT and Harris Corner Detection, coupled with k-NN, Random Forest, and SVM classifiers, are evaluated against modern CNN architectures including AlexNet, VGG-16, and ResNet50. The assessment focuses on accuracy, processing time, robustness, and scalability within robotic applications using the Fashion-MNIST and CIFAR-10 datasets. Furthermore, the study delves into the broader implications of deep learning in robotics, highlighting the potential enhancements in autonomous navigation, object manipulation, human-robot interaction, and more. Challenges such as data dependency, model interpretability, and generalization are also discussed, underlining the dual aspects of opportunities and obstacles presented by deep learning in the robotic domain. Code of this work can be found at https://github.com/ziyicheng427/A-Comparative-Analysis-of-Traditional-and-CNN-based-Object-Recognition-Techniques-in-Robotics.