一种检测范围自适应的目标跟踪算法

为处理目标的消失重现、形变及环境变化等问题, 要求跟踪算法有一定的检测与学习能力. 针对全局检测方法因冗余检测而造成检测效率低下的问题, 在基于P-N学习的跟踪框架的基础上, 提出一种自适应生成检测范围的目标跟踪算法. 通过引入卡尔曼滤波器(Kalman filter)对目标位置、尺度以及两者的变化速度进行预估, 在检测前根据预估信息自适应生成检测范围, 提高检测效率. 在公开的CoGD数据集上进行实验, 结果证明该算法较原始算法在准确度基本不变的基础上, 速度得到显著改善.

An Adaptive Approach for Generating Detective Range in Visual Object Tracking

Visual object tracking is a highly challenging task due to the fact that it suffers from many intrinsic problems, such as object disappearance, reappearance, deformation and environment variation. A tracking algorithm is required to have the ability of detecting and learning. Given the fact that the global detector would generate certain redundant detections, we propose an approach that adaptively generates the detective range for visual object tracking based on the framework of P-N learning. We introduce a Kalman filter to predict the location, the object scale as well as their changing rate. Then we utilize the predicted information to reduce the scope of detecting. Our experimental results on the pubic CoGD dataset show that our method increases the speed dramatically while without compromising the accuracy.