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。

Focusing on the issue that spatially regularized discriminative correlation filter (SRDCF) tracking algorithm has poor performance in handling rotation, out of view and heavy occlusion, we propose a visual tracking approach based on adaptive convolutional features. First, based on the principal component analysis of conv3-4 layer features in the VGG-NET model, the dimension of conv3-4 layer features is reduced from 256 to 130 by adaptive dimension reduction technique. Then, we maximize classifier score in the detection area and get the location and scale of target. In order to redetect the target in the case of tracking failure and achieve long-term tracking, we compare the confidence of the location with maximum score and train an online support vector machine (SVM) classifier. Finally, the tracking model is updated by the reliable tracking results which are determined by peak-to-sidelobe ratio. To verify the feasibility of the proposed algorithm, the results are compared with those obtained by thirty-eight kinds of tracking algorithms in one hundred video sequences of OTB-2015 benchmark. Experimental results indicate that the precision and success rate are respectively 0.804 and 0.607. The proposed approach has a ranking of one. Compared with SRDCF tracking algorithm, the proposed approach improves the precision and the success rate by 1.9% and 1.5%, respectively. In addition, the proposed approach is robust for rotation, out of view, heavy occlusion and other complex scenes.