基于多特征融合与ROI预测的红外目标跟踪算法

Infrared Target Tracking Algorithm Based on Multiple Feature Fusion and Region of Interest Prediction

针对当前红外目标检测与跟踪算法存在场景自适应能力弱、专用性强,以及在大视场条件下,首帧图像中小目标误检率高的问题,提出一种红外序列图像目标自适应阈值分割、检测与跟踪方法.选取目标移动速度、目标轮廓的面积和周长、以及自适应分割阈值与感兴趣区域位置为动态变量,建立动态决策准则.采用首帧目标检测算法计算出序列图像的第一帧图像目标的静态变量和部分动态变量,再采用改进的局部自适应阈值分割算法分割后续帧图像,然后利用静态与动态决策准则筛选出分割后的真实目标,最后计算并更新动态决策准则.红外靶标测试结果表明:该方法对不同场景具有较好的适应性,四个场景平均跟踪准确率为95.81%,微机平台平均每帧处理时间为10.93ms,嵌入式平台为26.79ms.

A method of adaptive threshold segmentation,detection and tracking for infrared sequence images is proposed to solve the problems that the current infrared target detection and tracking algorithm has weak scene adaptability,strong specificity and high false detection rate of small target in the first frame image under large field of view.The density,rectangularity and Hu invariant moments are selected as static variables to establish static decision criteria.The target moving speed,area and perimeter of target contour,adaptive segmentation threshold and location of ROI are selected as dynamic variables to establish dynamic decision criteria.The first frame target detection algorithm is used to calculate the target static variable and some of the dynamic features of the first frame image.The subsequent frame images are segmented by the improved local adaptive threshold segmentation algorithm and then the static and dynamic decision criteria are used to screen out the segmentation.Finally,the dynamic decision criteria are calculated and updated.The infrared target test results show that the method has good adaptability to different scenarios.By using this algorithm,the average tracking accuracy of the four scenarios is 95.81%,the average processing time per frame is 10.93 ms on microcomputer platform and 26.79 ms on embedded platform respectively.