Comparison of handcrafted features extraction techniques for traffic sign recognition

Traffic Sign Recognition (TSR) is a critical computer vision task for the advancement of Autonomous Driver Assistance Systems (ADAS) and autonomous vehicles, directly impacting road safety. While deep learning dominates current research, handcrafted feature extraction techniques remain relevant due to their interpretability, lower computational demands, and suitability for embedded systems. This paper presents a systematic empirical evaluation and comparison of five predominant handcrafted feature extraction methods: Histogram of Oriented Gradients (HOG), Scale-Invariant Feature Transform (SIFT), Bag of Words (BoW), Local Binary Patterns (LBP), and Gabor Filters for traffic sign recognition. The techniques are assessed on the German Traffic Sign Recognition Benchmark (GTSRB) using three classifiers: Support Vector Machine (SVM), Random Forest, and Decision Tree. Experimental results demonstrate that the choice of feature extractor and classifier significantly impacts performance, with HOG combined with an SVM classifier achieving the highest accuracy (95.2%). The study provides a clear performance hierarchy, revealing that HOG and LBP offer the best balance of accuracy and computational efficiency for this domain. We conclude that the optimal selection of a handcrafted feature extraction strategy is problem-dependent and provide concrete recommendations for implementing effective and efficient TSR systems.