Testing and Development of a Multi-Object Tracker Based on Deep Learning Techniques

In today’s scenario, computer vision is one of the fundamental research areas of artificial intelligence including object detection and object tracking which are the upcoming trends. In the present work, the TransTrack model has been reimplemented which includes ResNet50 as a backbone network. Further, to improve the overall performance of the TransTrack model, some recently developed backbones such as ResNet152 and SwinT have been incorporated. Moreover, to test the optimum training of the existing model, it has been retrained for higher epochs (such as 160 epochs). However, the accuracy of the existing tracker has improved significantly as the number of epochs increases but, at the same time, it increases the computational complexity. On the contrary, when the existing tracker (yields 68.4% accuracy with ResNet50) is incorporated with ResNet152 and SwinT, the performance has significantly improved (71% for ResNet152 and 71.6% for SwinT) by 1.7% and 2.1%, respectively. The mentioned analysis is enough to prove that the incorporation of a more efficient backbone has the potential to improve the performance of trackers. In the future, the newly introduced backbone may open the window to improve the performance of multi-object tracking (MOT).