An end to end trained hybrid CNN model for multi-object tracking

Abstract

A robust MOT (multi-object tracking) is very crucial for computer vision applications such as crowd density estimation and autonomous vehicles. Most of the existing mot approaches perform object tracking in a two task manner such as motion estimation and Re-identification but these approaches pose some drawbacks like the model is not end-to-end trained, the Re-Id required lots of identity switches thus incurred computational overhead and the performance further degrades in complex crowd scenarios. To overcome such drawbacks we are motivated to design an end-to-end trained DNN for MOT. The proposed model utilizes a matching technique that utilizes the relative scale between the boundary boxes and relative position calculates the relative distance between the objects for MOT. To solve the problems, we proposed a matching technique that poses two subtasks to efficiently scale up a single shot DNN tracking approach for an indefinite number of objects in the video frames. The proposed method uses a relative scale and relative position to matching between the detected and targeted objects. The achieved state-of-the-art results of the tasks allow to obtain high accuracy of tracking with detection and surpasses existing state-of-the-art methods by a huge margin on various public datasets. © 2021, The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature.