基于自适应卷积特征的目标跟踪算法

发布时间：2018-12-14 02:45

【摘要】：针对空间正则化相关滤波(SRDCF)跟踪算法在目标跟踪中旋转变化、超出视野和严重遮挡情况下存在跟踪失败的问题,提出了一种基于自适应卷积特征的目标跟踪算法。对VGG-Net模型中conv3-4层卷积特征进行主成分分析,利用自适应降维技术将conv3-4层特征维数由256维降至130维。在检测区域求取分类器最大响应位置及其目标尺度信息,并对最大响应位置的目标进行置信度比较,训练在线支持向量机(SVM)分类器,以便在跟踪失败的情况下,重新检测到目标而实现长期跟踪。计算跟踪位置的峰旁比,选取可靠跟踪结果,更新模型。采用OTB-2015评估基准的100组视频序列进行测试,并与38种跟踪方法进行对比,验证了本文算法的有效性。实验结果表明:本文算法跟踪精度为0.804,成功率为0.607,排名第一,与SRDCF算法相比,两者分别提高了1.9%和1.5%。针对目标发生旋转变化、超出视野和严重遮挡等复杂情况,本文算法均具有较强的稳健性。
[Abstract]:In order to solve the problem that the spatial regularization correlation filtering (SRDCF) tracking algorithm has the problem of tracking failure in the case of out-of-sight and severe occlusion, a target tracking algorithm based on adaptive convolution feature is proposed. Based on the principal component analysis of the convolution feature of conv3-4 layer in VGG-Net model, the characteristic dimension of conv3-4 layer is reduced from 256-dimension to 130D by adaptive dimensionality reduction technique. The maximum response position and its target scale information of the classifier are obtained in the detection region, and the confidence degree of the target in the maximum response position is compared. The online support vector machine (SVM) classifier is trained in case of tracking failure. Re-detect the target for long-term tracking. The peak to side ratio of tracking position is calculated, the reliable tracking results are selected, and the model is updated. The effectiveness of the proposed algorithm is verified by using 100 sets of video sequences evaluated by OTB-2015 and compared with 38 tracking methods. The experimental results show that the tracking accuracy of this algorithm is 0.804 and the success rate is 0.607, ranking first. Compared with SRDCF algorithm, the two algorithms are improved by 1.9% and 1.5% respectively. This algorithm is robust for complex situations such as rotation, out-of-field and severe occlusion.
【作者单位】：河北工业大学控制科学与工程学院;
【基金】：国家自然科学基金(61203076) 天津市自然科学基金(13JCQNJC03500)
【分类号】：TP391.41

【相似文献】