IPNet:帧间预测校验的多域卷积神经网络目标跟踪算法

针对现有卷积神经网络分类预测MDNet算法在线更新机制容易将错误样本引入网络模型,导致跟踪算法失效的问题,提出一种基于帧间预测校验的MDNet目标跟踪改进算法IPNet。该算法运用视频压缩领域的帧间预测方法和聚类算法,在前一帧目标跟踪位置的基础上,计算前后连续两帧中目标的相似度,估计出目标下一帧可能出现的候选区域,实现目标位置预测,达到校验跟踪结果的目的。IPNet算法能有效减少更新样本导致的跟踪失效问题,改善了在目标旋转、快速运动以及背景混杂等情况下的跟踪效果,提升了算法的跟踪性能。

IPNet:Interframe Prediction Verification Multi-domain Convolutional Neural Network Target Tracking Algorithm

The online update mechanism of the MDNet algorithm, the existing convolutional neural network classification and prediction algorithm, is easy to introduce error samples into the network model, which leads to the failure of the tracking algorithm. Aiming at this problem, this paper proposes a multi-domain convolutional neural network target tracking algorithm based on interframe prediction. The algorithm uses the interframe prediction method which is often used in the field of video compression and the clustering algorithm. Based on the target tracking position of the previous frame, the similarity of the targets in two consecutive frames is calculated, and the candidate regions that may appear in the next frame of the target are estimated. So the target position prediction is realized to verify the purpose of the tracking result. The algorithm can effectively reduce the tracking failure caused by updating samples, improve the tracking effect under the conditions of target rotation, fast motion and background aliasing, and improve the tracking performance of the algorithm.