CAJ | 학술논문

为满足光电跟踪系统图像跟踪抗遮挡、实时性的要求，提出了一种改进检测器和目标模型更新策略的 TLD 算法。首先，通过帧差法获得差分图像序列，其次，利用动态 Otsu 阈值对差分图像进行二值化处理，滤除背景差分像素，获取移动物体边界框，最后，产生局部滑动窗口，进行随机厥分类和最近邻分类；并且优化了目标模型更新策略。实验表明，对于分辨率为320×240的视频，改进算法较原算法跟踪速度提升比平均为1.50，满足系统的实时性要求；改进算法抗遮挡性能及在低对比度环境中的跟踪性能优于 Mean-Shift 算法，满足系统的抗遮挡要求。
위만족광전근종계통도상근종항차당、실시성적요구，제출료일충개진검측기화목표모형경신책략적 TLD 산법。수선，통과정차법획득차분도상서렬，기차，이용동태 Otsu 역치대차분도상진행이치화처리，려제배경차분상소，획취이동물체변계광，최후，산생국부활동창구，진행수궤궐분류화최근린분류；병차우화료목표모형경신책략。실험표명，대우분변솔위320×240적시빈，개진산법교원산법근종속도제승비평균위1.50，만족계통적실시성요구；개진산법항차당성능급재저대비도배경중적근종성능우우 Mean-Shift 산법，만족계통적항차당요구。
To meet the requirements of real-time and anti-occlusion tracking in optoelectronic tracking systems, an improved Tracking-Learning-Detection (TLD) approach with a modified detector and a modified strategy of module updating is proposed. Firstly, difference images are obtained by the difference between consecutive frames. Then, binary images are obtained by an adaptive Otsu threshold. As a result, difference pixels of background part are filtered and moving object bounding boxes are obtained from the binary images. Finally, local scanning-windows are generated for Ensemble Classifier and Nearest Neighbor Classifier. And the strategy of module updating is improved. The experimental results indicate that the average speedup per frame of the improved TLD approach reaches up to 1.50. Compared with the traditional TLD, the improved approach can meet system’s real-time requirement tracking performance of the improved TLD approach under occlusions or in low contrast environment is better than the Mean-Shift approach, and the improved approach can meet system’s requirement of anti-occlusion tracking.