面向目标检测应用场景的运动相机阵列研究与实现

DEVELOPMENT OF MOBILE CAMERA ARRAYS FOR OBJECT DETECTION APPLICATIONS

杨子江

11749247

硕士

计算机科学与技术

郝祁

2020-06-02

2020-07-08

哈尔滨工业大学

深圳

智能相机网络在公共安全、室内异常检测场景中有着广泛应用。近年来，将可配置参数的相机与板载处理器和无线通信模块结合进而构建检测系统的方式逐渐成为主流技术，对此类系统的拓扑结构、智能资源分配和重配置算法的设计吸引了许多研究关注。主要技术挑战包括：（1）如何提高系统的可扩展性和稳定性；（2）如何充分利用多视角信息，并有效地执行多相机参数重新配置以完成预期任务；（3）如何验证系统部署和提出的算法的有效性。本文提出了一种新颖的智能运动相机阵列系统集成框架。这项工作的优势体现在三个方面：（1）使用分布式拓扑保证系统的可扩展性，并依靠惯性传感器和里程计数据反馈结合经典控制策略，提高阵列单元的控制精度以及高机动性下的运动稳定性；（2）提出了两种启发式搜索方法来规划系统中相机的运动，完成参数重配置，并实现了基于深度学习的图像处理算法，以帮助提高系统的多视角检测和场景分析性能；（3）相比现有工作，使用Unreal Engine搭建的3D仿真环境与硬件平台之间算法迁移性强，仿真结果可以和实验一体化，有利于安全高效地开展在不同约束条件下对部署运动相机阵列和优化系统性能的研究。根据提出的系统框架，搭建了对应的硬件实验平台和3D仿真平台，并依托上述平台进行了大量实验。实验结果表明，在目标检测相关任务中，与现有其他方法相比，启发式重配置方法可以在较短的时间内取得良好的效果；并且整个系统工作时运行稳定，可以满足各种性能标准，为扩展高级应用奠定了坚实的基础。

The smart camera networks (SCNs) have been widely used in various scenarios such as public safety and indoor anomaly detection. In recent years, combining reconfigurable cameras and on-board processing units with wireless communication modules to build surveillance systems has become a mainstream technology. Developments of advanced topologies, intelligent resource allocation and parameter reconfiguration schemes for SCNs have attracted many research attentions. The related major technical challenges include: (1) how to increase system scalability and stability; (2) how to make use of multi-view information and efficiently perform multi-camera parameter reconfiguration to carry out expected tasks, and (3) how to verify the effectiveness of system deployments and proposed algorithms.This work proposes a novel integrated framework for the development of smart mobile camera array systems. Advantages of this work include three aspects: (1) a distributed topology is used to achieve the system scalability, and each individual unit's control accuracy and stability under high motility are improved by applying classic control methods to the feedback data from IMU and odometry; (2) heuristic search methods are developed to achieve system motion planning for multi-view based tasks, and deep learning based image processing algorithms are implemented to help improve the multi-view detection and scene analysis performances; (3) compared with existing works, algorithm migrations between the Unreal Engine based 3D simulation environment and the hardware platform are flexible, so results of experiments and simulations can be integrated, which is conducive to the safe and efficient studies of deploying mobile camera arrays and optimizing system performance under different constraints.Corresponding hardware and 3D simulation platforms are built and numerous experiments are carried out. Results show that in terms of target detection tasks, proposed heuristic reconfiguration methods can achieve good results in a shorter time compared with other existing methods; and the whole system works stably to meet various performance requirements, laying a solid foundation for extending advanced applications.

运动相机阵列 3D仿真 运动规划 图像处理 系统设计

mobile camera arrays 3D simulation motion planning image processing system design

中文

联合培养

南方科技大学

杨子江. 面向目标检测应用场景的运动相机阵列研究与实现[D]. 深圳. 哈尔滨工业大学,2020.