Adaptive Search Area in the Object Tracking [Nesne Takibinde Uyarlanabilir Arama Alani]

Abstract

Object tracking is the automatic determining of position and size information of the target object in a video along the frames. Object tracking methods generally use a search area in a smaller size, rather than using an entire frame view to extract position and size information of any frame target object. This search area is usually chosen so that the object position in the previous frame is centered and will be 'k' times the object size. In the general, after the object position in the frame is determined, the relevant object in the search area is taken as positive and the remaining part is taken as negative and classifier updates are made. In this way, the object can be strong against the appearance changes throughout the frame. The size of the search area in the frame directly affects object tracking performance and speed performance. The oversized selection of the search area can reduce speed performance on object tracking while allowing a more powerful classifier to be trained. Likewise, if the search area is selected too small, then the object will run at very high speeds. In this study, an adaptive search area (UAA) algorithm is proposed, which can adaptive select the search area used for object tracking. The developed UAA algorithm is applied to HCFT (Hierarchical Convolutional Features for Visual Tracking) method for the determine the effects on tracking performance. Then, practical tests were conducted to compare with the original HCFT method. As a result of the tests, UAA method has been found to give very strong results in terms of speed and performance. With the UAA algorithm developed, object detection can be performed at higher speeds, keeping the search area small in video scenarios. On the other hand, in difficult scenarios, the object tracking accuracy is improved by keeping the search area larger. Thanks to the improved UAA algorithm, it has made object tracking at higher speeds by reducing the search area in video scenarios in which the object detection is simple. On the other hand, in difficult scenarios, the object tracking accuracy is improved by keeping the search area larger. © 2018 IEEE.