Local track to detect for video object detection. (12th March 2022)

• Record Type: Journal Article
• Authors: Zeng, Biao; Zhong, Shan; Zhou, Lifan; Wang, Zhaohui; Gong, Shengrong
• Is Part Of: International journal of computer applications technology. Volume 67:Number 2/3(2021)
• Page Start: 119
• Page End: 128
• Publication Date: 2022-03-12
• Subjects: video object detection -- local detection -- detect and track -- movement prediction -- efficient detection -- CNN; Technology -- Data processing -- Periodicals; 620.00285
• Languages: English
• ISSNs: 0952-8091

Abstract

The existing methods for video object detection are generally achieved from searching the objects through the entire image. However, they always suffer from large computation consumption as a result of dozens of similar images needing to be operated. To relieve this problem, we propose a Local Track to Detect (LTD) framework to detect video objects by predicting the movements of objects in local areas. LTD can automatically determine key frames and non-key frames, the objects in key frames can be detected by the single frame detector, and the objects in non-key frames can be efficiently detected by the movement prediction module. LTD also has a Siamese module to predict whether objects between the key frame and the non-key frame are the same object to ensure the accuracy of the movement prediction module. Compared with other previous work, our method is more efficient and achieves state-of-the-art performance.