Particle Filter based Single Shot MultiBox Detector for Human Moving Prediction

Abstract

Moving object tracking has become a center of attention for computer vision researchers. It is quite challenging to track a moving object correctly, especially when the object has occlusion, changes in illumination, unexpected movements, and arbitrary poses. To enhance the accuracy of the moving object detector, we proposed SSD (Single Shot Multibox Detector) in addition to PF (Particle Filter) to provide prediction of moving human. Performance evaluation was done with the comparison to the previously proposed HOG-SVM as a detector. The proposed system has been successfully tested in two videos. PF based SSD with 100 particles performs well, with RMSE 7.44 and 91.24 effective particle. The results show that the addition of SSD in measurement process could enhance the PF's performance to track moving human. The results have also shown that the proposed method was successfully implemented in combination with a specific color detection to track a specific human object.