Improved mask RCNN and cosine similarity using RGBD segmentation for Occlusion handling in Multi Object Tracking

In this study, additional depth images were used to enrich the information in each image pixel. Segmentation, by its nature capable process up pixel level. So, it can detect smallest part of object, even when it’s overlapped with another object. By using segmentation, main goal is be able maintain tracking longer object starts occluded until severely right before completely disappeared. Object based on detection was developed modifying Mask R-CNN architecture RGBD images. The results feature extracted HOG, and them got compared target objects. comparison cosine similarity calculation, maximum value detected would update for next frame. evaluation model mAP calculation. late fusion had a higher 5% than RGB. It 68,234% 63,668%, respectively. Meanwhile, uses traditional method calculating id switching during process. Out 295 frames, original ten ID times. On other hand, proposed much better ids close 0. Keywords—Occlusion, RGBD, R-CNN, Late fusion, Cosine