Survey and Experimental Comparison of Machine Learning Models for Motorcycle Detection

Road accidents are largely related to motorcycles and many cases resulted in severe injuries and fatalities. Automatic detection of motorcycles is considered a useful measure for monitoring evaluating predicting and eventually planning to prevent further motorcycle-related accidents. Many researchers proposed techniques to detect and track vehicles but only few focused on detecting motorcycles. In this paper we reviewed and conducted experiments on 20 different machine learning models with different sizes and views (front rear top) of traffic images to evaluate motorcycle detection's accuracy and time. By setting 60% threshold of accuracy on images of 480⨯270 pixels the best detection model is rfcn_resnet101_coco which provided average accuracy detection time and efficiency at 65.85% 1.91 frames/second and 34.55 respectively.