Special Pedestrian and Head Pose Detection for Autonomous Vehicles

Pedestrian safety is a major concern due to increased traffic density and due to driver-pedestrian errors. ADAS (Advanced Driving Assistance Systems) should take different actions for special pedestrians like wheelchair pedestrian. Not much work is done for detecting special pedestrian. In this paper we propose cost effective real-time model for special pedestrian detection and head pose detection. We first detect wheelchair pedestrian and then predict head pose for the Region of Interest (ROI). It is observed that accuracy of wheelchair detection increases considerably by using head pose detection. Our model employs an ensemble of algorithms based on Histogram of Gradient (HOG) for pedestrian features, Modified Census Transform (MCT) for head features and head pose features and SVM for classification. We compare the proposed model against transfer learning approach based on Inception-v3 model.