针对车辆行进中, 红外热像仪拍摄的视频序列存在复杂的随机抖动, 提出基于BRISK特征点匹配的运动估计算法, 计算出高精度全局运动矢量, 同时对于特征点匹配时出现误匹配及场景中存在前景运动物体的情况, 采用模糊聚类法分离全局运动和局部运动, 提高了算法的鲁棒性。提出了基于Kalman粒子滤波算法, 有效实现了复杂扫描运动和随机抖动的分离, 并利用双线性插值法进行图像补偿。采用快速图像拼接法进行“未定义区域”处理, 实现了图像全景输出。还利用车载红外热像仪实际拍摄的红外视频进行了稳像实验。实验结果表明, 视频序列获得了很好的稳像效果, 能够满足实际应用要求。

Focusing on the random vibration phenomena in infrared video sequences taken from moving vehicles, a BRISK characteristic point matching based high-precision global motion estimation algorithm is proposed. For mismatch in feature point matching and moving objects in the scene, fuzzy clustering method is used to separate global motion and local motion, which improves the robustness of the algorithm. A Kalman particle filter is used to separate the complex scanning movements and random vibration effectively. Meanwhile, we use bilinear interpolation for image compensation. Finally, a panoramic image is acquired by processing the “undefined region” with fast image alignment. A series of experiments have been carried out using video sequences taken on vehicle-borne thermal camera and the experimental results illustrate that the video sequences are well stabilized, which meet the requirements for practical use.