Analysis of Kernel Performance in Support Vector Machine Using Seven Features Extraction for Obstacle Detection

Many electric powered wheelchairs (EPW) users fall due to the user’s carelessness of the road conditions in front of them that will have a significant impact on accidents. The process for detecting road conditions is one solution to maintain the safety of EPW users. This research is conducted to develop autonomous systems in the wheelchair to detect stair descent and floor obstacles. The system accomplished to prevent fatal risks occurs to the user, such as falling from the stairs that cause fractures. Moreover, the main goal of the system expansion is to identify the best kernel class from the support vector machine (SVM) classification method to distinguish the stair descent and the floor. This experiment is completed using the SVM method classified into four kernel functions: linear, polynomial, Gaussian, and Sigmoid kernel class, and also associated with gray-level co-occurrence matrix (GLCM) features extraction. The SVM produces the best result for detecting used linear kernel function with GLCM parameters (d = 1, θ = 0) was reached an average of accuracy is 89.0% for image data testing and video testing is 82.6%.