红外弱小多目标实时处理

发布时间：2019-06-05 05:47

【摘要】：红外成像技术隐蔽性好、抗干扰能力强且能全天候工作,被广泛应用于民用和军事领域。但红外弱小目标的成像距离远,背景杂波干扰严重,使得其成像信噪比低、结构信息不足,对红外图像预处理方法、弱小目标检测以及弱小目标跟踪方法的研究成为红外弱小多目标实时处理技术的关键,在红外制导等领域起到关键作用。本论文以实际科研项目为研究背景,基于FPGA+DSP架构实现红外图像的实时处理方法的设计与优化。红外弱小多目标图像背景杂波和噪声干扰严重,图像预处理至关重要,于是本文采用改进中值滤波来适应不同噪声类型和噪声密度,在改变滑动窗口长度的同时改变窗口形状,并保证以较小的窗口进行滤波。基于目标和背景特征,采用改进形态学背景抑制算法,通过尺度变化的半圈组合型结构元提取起伏变化的图像背景,实现背景杂波的有效抑制。在弱小多目标检测方面,结合带尺度因子的点扩散模型,在拉普拉斯高斯尺度空间下表示图像,初步确定目标位置和大小,通过各向差异度均值的阈值判断提取出所有可疑目标,再根据目标大小结合各向差异度实现真实目标的检测。在弱小多目标的跟踪方面,为了实现可靠跟踪,匹配不同数据更新速率的滤波器,对低速运动目标匹配均值漂移卡尔曼滤波器,而高机动运动目标匹配改进均值漂移粒子滤波器,并进行交互式融合,得到目标跟踪结果,而为了实现多目标的可靠跟踪,通过结合马尔可夫随机网络,考虑了各个目标的相邻目标的状态,估计出每个目标最大联合后验概率,更新滤波器参数和粒子权值,进行多目标位置估计。基于红外实时图像处理平台,通过算法移植,对分辨率为640*512的中波红外相机连续拍摄的大量数据进行实时处理,处理结果表明,在图像预处理方面,本文采用的改进中值滤波噪声平滑算法和改进形态学背景抑制算法具有较好的处理效果和实时性;在多目标的检测方面,本文采用的基于尺度空间的各向差异度检测算法相比已有算法,鲁棒性更好,具有更高的检测率,同时单帧的平均处理时间小于5ms,满足处理实时性要求;在多目标的跟踪方面,对于本文采用的结合马尔可夫随机网络的多模型改进卡尔曼粒子滤波方法,其跟踪正确率是传统交互式多模型算法的3倍,处理速度能达到72帧/S,跟踪可靠性高、实时性好。综合而言,本文的红外弱小多目标实时处理方法可靠,有很高的实际应用价值。
[Abstract]:Infrared imaging technology is widely used in civil and military fields because of its good concealment, strong anti-interference ability and all-weather work. However, the imaging distance of infrared small and weak targets is long and the background clutter interference is serious, which makes the imaging signal-to-noise ratio (SNR) low and the structure information insufficient. The research of weak and small target detection and weak small target tracking method has become the key to infrared small and weak multi-target real-time processing technology, and plays a key role in infrared guidance and other fields. This paper takes the actual scientific research project as the research background, based on FPGA DSP architecture to realize the design and optimization of infrared image real-time processing method. The background clutter and noise interference of infrared weak and small multi-target image is serious, so image preprocessing is very important. Therefore, this paper adopts improved median filtering to adapt to different noise types and noise density, and changes the window shape while changing the length of sliding window. And make sure to filter with a smaller window. Based on the target and background characteristics, the improved morphological background suppression algorithm is adopted, and the fluctuating image background is extracted by the semi-circular combined structure element with scale change, and the background clutter is effectively suppressed. In the aspect of weak and small multi-target detection, combined with the point diffusion model with scale factor, the image is represented in Laplace Gao Si scale space, and the target position and size are preliminarily determined. All suspicious targets are extracted by the threshold judgment of the mean value of each direction difference, and then the real target detection is realized according to the target size combined with the difference degree of each direction. In the aspect of weak and small multi-target tracking, in order to realize reliable tracking and match the filter with different data update rate, the mean drift Kalman filter is matched to the low-speed moving target. The high maneuvering moving target matching improves the mean drift particle filter, and carries on the interactive fusion to obtain the target tracking result. In order to realize the reliable tracking of the multi-target, in order to realize the reliable tracking of the multi-target, by combining the Markov random network, Considering the state of the adjacent targets of each target, the maximum joint posterior probability of each target is estimated, the filter parameters and particle weights are updated, and the multi-target position estimation is carried out. Based on the infrared real-time image processing platform, a large number of data taken continuously by the medium-wave infrared camera with a resolution of 640 / 512 are processed in real time through the algorithm transplantation. The processing results show that in the aspect of image preprocessing, The improved median filter noise smoothing algorithm and the improved morphological background suppression algorithm used in this paper have good processing effect and real-time performance. In the aspect of multi-target detection, the proposed algorithm based on scale space has better robustness and higher detection rate than the existing algorithms, and the average processing time of a single frame is less than 5 Ms. Meet the real-time requirements of processing; In the aspect of multi-target tracking, the tracking accuracy of the multi-model improved Kalman particle filter combined with Markov stochastic network is three times higher than that of the traditional interactive multi-model algorithm, and the processing speed can reach 72 frames / S. The tracking reliability is high and the real-time performance is good. To sum up, the real-time processing method of infrared weak and small multi-target in this paper is reliable and has high practical application value.
【学位授予单位】：苏州科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】