Multiple Object Tracking by Bounding Boxes Without Using Texture Information and Optical Flow

Object tracking is a key task in many applications using video analytics. While there huge number of algo-rithms to track objects, still need for new methods solve the correspondence problem under certaincircumstances. In our article, we assume very typical but open scenario: image object detector hasalready identified objects be tracked; thus, have labels, confidence values, and bounding boxes ineach frame captured at low sampling rate. That is, optical flow difficult applied (also dueto bad lighting conditions, cluttered or homogeneous areas strong ego-motion), moreover, objectslook similar (having same category labels). Our proposed approach based on Hungarian method andincorporates above information into cost function evaluating possible pairings objects. To considerthe uncertainty detector, elements confusion matrix also contribute pairs, as wellas probability spatial translations prior observations. As use case, apply algorithm adata-set, where images were from onboard cameras traffic signs detected by RetinaNet. Weanalyze performance with different parameter settings.