PROJECT TITLE :

Deformable Object Tracking with Gated Fusion

ABSTRACT:

Integration of tracking-by-detection frameworks with convolutional neural networks has received increasing attention (CNNs). Existing detection-based tracking approaches, on the other hand, are unable to keep track of items that exhibit extreme fluctuation in appearance. As a result, typical convolutional operations may not be able to identify the correct answer when the object's position or ambient conditions change. For the tracking-by-detection architecture, we offer a deformable convolution layer that enhances the target appearance representations Deformable convolution, which adaptively enriches the original properties of the target, is our goal. The deformable convolution also captures fluctuations in the original look, and we suggest a gated fusion technique to regulate this. Deformable convolution enhances the CNN classifier's ability to distinguish between the target object and its backdrop. There is no doubt that the suggested tracker is superior than current approaches, as evidenced by rigorous testing on industry-standard benchmarks.