CNN and HOG based comparison study for complete occlusion handling in human tracking

Abstract

Applications in the Human Machine Interaction (HMI) field are increasing since people and robots share the same environment and perform activities together. Therefore, robust detection and strong tracking methods are required. This study is about human detection and tracking. For detection, two different approaches, Histogram of Oriented Gradients (HOG)-Support Vector Machine (SVM) and Convolutional Neural Network (CNN) are used. The video used in the application is recorded in a noisy environment, and in some frames, the human is completely covered by other objects in the environment, i.e. exposed to complete occlusion. For this reason, Kalman Filter (KF) and Particle Filter (PF) based tracking is performed. Different from these two traditional filters, a hybrid Kalman-Particle Filter (KPF) has also been proposed. The contribution of this study is to compare CNN with HOG-SVM, which is described as the most successful human detection method. As a result of the study, it was found that CNN provides successful results especially in case of occlusion. This study is also important in terms of proposing a hybrid KPF and comparing the filters in the case of complete occlusion. The results showed that for human tracking, CNN using KF performed better performance throughout the video. The proposed KPF also outperformed PF and this superiority became much more prominent in the case of complete occlusion.