基于深度学习的目标抗干扰跟踪算法

摘要

Aimed at the interference of similar background in the target tracking algorithm and rotation of the target frame, the detection scheme of a single shot MultiBox detector (SSD) is proposed, which effectively avoids drifting of the tracking box. First, pre-training of the depth learning model is carried out on a specific kind of target to be tracked, and target location and tracking are completed by using the fusion discriminant scale space algorithm, which is designed in this paper. Discriminant scale spatial model is used to locate the target; feature detection is carried out in the candidate region; a kind of motion estimation and elimination system is designed to ensure uniqueness of the target of the candidate region:and finally the precise positioning of the target is established. Experiments show that this method can effectively avoid the tracing of the tracking frame, caused by similar background interference and occlusion, and can robustly track a fast-moving target and changes in the scale and shape.%针对目标跟踪算法中相似背景的干扰及目标自身旋转导致跟踪框漂移的情况,提出一种融入SSD(Single Shot MultiBox Detecter)检测的方案,从而有效地避免了跟踪框的漂移.首先对要跟踪的特定种类的目标进行深度学习检测模型的预训练,然后利用本文所设计的融合判别尺度空间算法完成目标定位和跟踪.由判别尺度空间模型对目标实施初步定位,在候选区域进行特征检测,并设计了一种运动估计淘汰体制,以保证候选区域目标的唯一性,最终完成目标的精确定位.实验证明,该方法能有效避免相似背景干扰和遮挡时所造成的跟踪框漂移,同时在目标快速运动,尺度和形状变化时均能完成鲁棒性的跟踪.