Multi-view data fusion in multi-object tracking with probability density-based ordered weighted aggregation

In this paper, a method is presented for Multi-Object Tracking (MOT) in presence of partial or complete occlusions. This work focuses on improved object detection and data association in a single view, and also fuses data from multiple views using the Ordered Weighted Aggregation (OWA) algorithm. Hence, a deep learning model was proposed to detect objects more accurately in a tracking-by-detection framework. This paper aims to enhance object detection, data association, and the trajectories of the objects in the MOT algorithm respectively by applying Mask R-CNN, Zernike Moments and combination of several similarity metrics, and fusion of multi-camera information by probability density-based OWA (PD-OWA). The inter-frame detected objects are matched based on the appropriate similarity metrics. In the fusion part, the Kernel Density Estimation (KDE) is utilized to assign weight coefficient to each camera view and determine the descending order of data in the OWA algorithm. Finally, the positions of each object coming from different views are weighted and aggregated. The results show that the proposed method improves object detection, association performance, and tracking trajectory in the "PETS09-S2L1'' and the "EPFL Terrace" video sequences and achieved 81.6% and 79.6% multiple objects tracking accuracy (MOTA), respectively.