遥感图像上颜色增强旋转不变霍夫森林目标检测方法的并行计算研究

Color-enhanced Rotation-Invariant Hough Forest (CRIHF) is a stable, efficient and versatile object detection method for remotely sensed imagery. However it is often forbiddingly slow to be successfully applied in many practical applications. Meanwhile, with the development of semi-conductor technology down to the atomic level, further development of computing hardware has required the use of parallel infra-structure. This new trend of development has often rendered algorithmic design a bottleneck as most complex algorithms are traditionally designed for the sequential computer model. Historically, parallel computation has been used to significantly improve the speed of certain scientific computation problems such as matrix computation and scientific simulations due to their structural regularity and excellent scalability. However, designing efficient algorithms for more complex problems to better utilize parallel hardware has been always a key research question in scientific computation and it is becoming important due to the wide-spread use of parallel hardware. Following such tendency, this project proposes to study the method and application of combining parallel computation and the CRIHF method. The main research contents include: (1) Design of a parallel computation model of the CRIHF method and its implementation using multiple types of parallel infra-structure. (2) Experiments of the parallel algorithm in high-throughput applications scenarios including real-time or near real time detection in a satellite image processing center and on unmanned aerial vehicles for verification of the parallel algorithm. (3) In-depth study on the influence of important parameters of the CRIHF to provide better guide for non-parallel implementations for practical use and developing extensions of the CRIHF method that are potentially time-consuming and can only be handled by parallel computing. By the study of the project, problems of CRIHF mentioned can be expected to be solved. The application range of CRIHF method can be expected to extended, which will provide a new fast object detection method on remotely sensed imagery which can be successfully applied in daily operations.

颜色增强旋转不变霍夫森林方法（CRIHF）适合于遥感图像目标检测，具有稳健、有效、通用性好的特点，然而此方法还存在检测速度慢，以及由此带来的高耗时特性较难研究、较难在工程应用中推广的问题。随着半导体技术发展到接近原子尺度，计算硬件需采用并行结构继续发展。并行计算因其优良的可扩展性以及充分利用并行计算硬件的能力可以大幅提高科学计算速度，正在成为科学计算的主流趋势。本项目顺应这一潮流，研究CRIHF方法并行计算的内在模型和外在应用，主要研究内容包括：（1）CRIHF方法在多种并行架构下的并行计算模型及其工程实现，（2）以卫星数据处理和无人机在航目标检测为应用场景验证方法，并总结重要参数对准确度、速度的影响规律。（3）在并行计算的支持下研究方法的高耗时特性扩展。通过本项目的研究，可以解决CRIHF方法的上述难题，扩展方法的应用范围，为快速遥感图像目标检测提供新方法。

本项目以遥感影像中旋转不变颜色增强纹元森林目标检测方法为起点，着眼于提升计算密集遥感处理任务的效率，重点研究了随机森林并行化、遥感应用背景下该方法的服务器端并行化和移动终端并行化，服务器端并行化在一定条件下取得了10倍左右的速度增益，在移动端也实现了相应的方法并达到了一倍的速度增益；结合项目研究发展遥感目标检测方法以及并行快速算法的宗旨，在项目框架内扩展研究了随机森林的多模态扩展、基于视频数据的运动暗弱目标检测与跟踪、以及位置分配模型VAOMP的并行化启发式算法，在相关领域获得了一系列创新性成果，到项目结题时在IEEE Transactions on Geoscience and Remote Sensing, IEEE Transactions on Aerospace and Electronic Systems, International Journal of Geographical Information Science等高水平国际刊物上已发表或接受了4篇相关SCI论文。