科研之友
微信
新浪微博
Facebook
分享链接
摘要
Target tracking is an important branch of computer vision, which includes three stages: image sequence optimizing, target expressing and target detecting. The target detecting stage is an important factor that influences the tracking performance. Therefore, how to obtain a more accurate and robust target detecting method becomes an urgent problem. Online sequential extreme learning machine (OSELM) is a kind of online learning method based on extreme learning machine. OSELM completes incremental learning by combining with the existed model when dynamic training samples are arriving. That OSELM has advantages including fast-speed and incremental learning suggests that is suitable for target detecting. Nevertheless, the target detecting process is different from the traditional classification for two causes: (1) target detecting is the dynamic process in that the position and rotation of the target are changing with time, and therefore the original OSELM method fails to obtain the most optimal target object from classified samples, (2) the tracking result frame depends on the previous frame, thus if the noisy sample is used as the target object, it would generates an impact to the tracking performance. To alleviate above-mentioned problems, this paper proposes an interesting and efficient target tracking method based on OSELM. In this method, we obtain the appropriate target object by judging the position relationship between each classified sample and the classification boundary. Moreover, we develop a kind of method that is similar to clustering to avoid tracking drift from noisy samples. The new target tracking method improves the performance remarkably, and eliminates the tracking drift from noisy samples. The proposed method is validated on six kinds of challenging image sequences.
- 出版日期2017-7
- 单位大连理工大学
