基于改进YOLOv3和核相关滤波算法的旋转弹目标探测算法

旋转弹的电视摄像头拍摄画面会产生旋转及抖动模糊，在预先侦查目标数据较少且末制导段视野目标较小的情况下，目标难以精确探测，为此提出一种基于改进YOLOv3和核相关滤波（KCF）算法的目标检测与跟踪算法，通过深度学习实现目标的自动检测。制作模拟山地打击场景的数据集，基于少量数据样本的前提，模拟不同天气、光照、运动及旋转模糊等复杂环境，完成在网络学习中数据的增强和扩充；通过在YOLOv3网络基础上添加Inception多尺度分支结构，增加网络对于目标不同尺寸的适应性，减少网络层数，更能适应对小目标的检测；在实现目标定位方法上，将目标检测与跟踪算法相融合，提出一种目标丢失判别机制，并利用弹道的速度—时间信息更新目标跟踪框尺度。仿真实验结果表明，相比原始算法，改进算法能更有效实现复杂环境下的目标检测和跟踪。

A Target Detecting Algorithm for Spinning Projectile Based on Improved YOLOv3 and KCF

The image captured by the spinning projectile-borne TV camerawill rotate and jitter to become blurry. It is difficult to detect a target accurately when the target data is less in advance detection and the field of view in the terminal guidance phase is small. A target detection and tracking algorithm based on improved YOLOv3 and kernelized correlation filter (KCF) is proposed.On the premise of a small number of data samples，the complex environments such as different weather，illumination，motion，and rotation blur are simulated to complete the data enhancement and expansion in network learning; By adding the multi-scale branch structure of Induction based on YOLOv3 network，the adaptability of the network to different sizes of targets is increased and the number of network layers is reduced for small target detection. In the realization of target location method， the target detection is combined with tracking algorithm，a target loss discrimination mechanism based on Gaussian threshold is proposed ，and the target frame scale is updated by using the velocity-time information of trajectory. Simulated results show that the improved algorithm can achieve the target detection and tracking in the complex environment more effectively.