Tracking Using Weighted Block Compressed Sensing and Location Prediction
-
摘要: 针对矩形跟踪框在边缘处包含较多背景信息的问题,该文提出一种基于规范化梯度特征的带权分块压缩感知的目标特征提取方法。该方法将压缩感知测量矩阵转化为分块对角矩阵,且根据块的重要程度分配适当的权重,缩小测量矩阵规模,简化特征提取运算,弱化背景干扰。然后将提取的特征输入变先验概率的贝叶斯分类器,变先验概率的分类器充分利用已有的跟踪结果,从一定程度预测了目标的运动方向,减小候选目标的分类歧义性,使得每一帧的分类函数根据以往跟踪结果进行变化,提高了分类的准确度。实验在8个具有常见跟踪难度的序列中测试,并与目前较流行的4种目标跟踪算法在跟踪效果、成功率等方面进行比较,结果从多个角度表明,该文提出的目标跟踪算法具有较高的准确度和稳定性。Abstract: To reduce side effects of background information included in the outer parts of tracking rectangular boxes, a weighted block compressed sensing feature extraction method is proposed based on normalized gradient features. The compressed sensing measurement matrix is converted to a block diagonal matrix. Appropriate weights are assigned to different blocks according to the importance of the blocks. It aims to reduce the measurement matrix size, weaken background interference and simplify feature extraction. Then the extracted features are inputted into Bayesian classifier with adaptive priori probabilities, which is proposed to make full use of existing tracking results. To some extent the classifier with variable priori probabilities can predict the direction of the moving targets, and reduce the ambiguities of target candidates. Each frame classification function changes according to the results of the previous track to improve the classification accuracy. In the experiments compared with four state-of-the-art tracking algorithms on 8 commonly used tracking test sequences, the proposed target tracking algorithm has higher accuracy and stability in terms of tracking results and success rate.
计量
- 文章访问数: 1275
- HTML全文浏览量: 130
- PDF下载量: 990
- 被引次数: 0