COMPARATIVE ANALYSIS OF FEATURE DESCRIPTORS FOR OBJECT RECOGNITION UNDER DIVERSE ENVIRONMENT

Abstract

Object detection is performed in several applications, especially in self-driving cars, where the exact detection of objects could reduce human error-related accidents to a large extent. Most traditional approaches to object detection involve the use of predefined features, and their practical use can be limited by low adaptability and robustness in various conditions. On the other hand, deep learning (DL) methods collect a considerable number of datasets to establish their personally productive, broad feature sets for deeper detection performance. The research presents a rigorous comparative analysis of feature descriptors for the object detection task across different environments, examining the domain of both traditional and DL based methods. Thus, the paper conducts an assessment of numerous feature descriptors, such as Scale-Invariant Feature Transform (SIFT), Speeded-Up Robust Features (SURF), Oriented FAST and Rotated BRIEF (ORB), as well as current DL models like Convolutional Neural Networks (CNN), YOLO, and Vision Transformers (ViTs). Utilizing a Support Vector Machine (SVM) for classification, the study rigorously evaluates the performance of each method employing the Pascal VOC 2007 dataset. The experimental analysis of the proposed model is conducted utilizing a Python tool. The results demonstrate that the ViTs + SVM combination significantly outperforms other models, achieving an impressive accuracy of 95.04%. This finding underscores the advantages of deep learning in effectively capturing complex patterns and enhancing the robustness of object detection systems.