基于无人机平台的目标检测与人机交互算法研究

发布时间：2018-08-28 11:37

【摘要】：近年来,随着无人机控制技术的日趋成熟和人工智能技术的蓬勃发展,多旋翼无人机除了执行传统的航拍、植保、战地侦查等任务之外,还被赋予了主动跟踪目标,自主目标识别等智能功能。目前,多旋翼无人机大部分的应用还是集中在民用娱乐方面,而无人机的人机交互方式仍主要依赖于较为复杂的遥控器控制,遥控器较大的体积妨碍了小型多旋翼无人机的便携性,遥控器控制的操纵不便性阻碍了无人机使用人群的推广。因此,需要设计一种不依赖于外接设备的智能人机交互方法。人体姿态信息较为明显且对人较为亲和,通过计算机视觉与人工智能技术,对人体姿态信息进行识别,就可以完成人机交互任务。首先,针对无人机在空中飞行时机械振动造成拍摄画面抖动的情况,采用中心区域模板匹配的方法,在算法上实现了对无人机机械振动的补偿,使得画面较为稳定并易于后续算法处理。接着,对图像中的人体目标进行检测。考虑使用了两种解决方案,分别用于地面站和机上处理情况。第一种是使用结构复杂,运算量大的卷积神经网络检测器Faster R-CNN完成候选区域的生成,目标种类的分类与人体目标位置的定位,该方法准确度高,能够适应复杂的背景环境,但运算量庞大,无法在机载嵌入式计算平台上实现实时性。另一种是使用前景信息提取的方法生成候选框,再使用结构较为简单的卷积神经网络完成分类任务,该方法需要先对场景信息进行建模,但运算量小,可以在机载嵌入式平台上实现实时性。然后,对检测到的人的姿态信息进行识别。为了确定检测到的人是否为需要进行人机交互的对象,设计了一个挥手动作检测器作为人体姿态信息检测器的开关,增强了系统的安全性。对通过了挥手动作检测的对象,设计了4种明显的姿势,无人机通过对姿势种类的识别完成人机交互任务。为了提高运行速度并充分利用已处理信息,将框选到的目标前景信息直接作为特征输入多层全连接神经网络进行识别。最后,将算法移植到了机载嵌入式设备上并进行了飞行实验验证。结果表明,系统能够在保持较高识别准确率的同时实现了机上处理的实时性,能够很好的完成人机交互任务。
[Abstract]:In recent years, with the increasing maturity of UAV control technology and the vigorous development of artificial intelligence technology, multi-rotors UAVs have been given active tracking targets in addition to performing traditional aerial photography, plant protection, field reconnaissance and other tasks. Autonomous target recognition and other intelligent functions. At present, most of the applications of multi-rotor UAV are concentrated in the field of civil entertainment, and the man-machine interaction mode of UAV still mainly depends on the more complex remote control. The large volume of remote control hinders the portability of small multi-rotor UAV, and the inconvenience of remote control hinders the popularizing of UAV users. Therefore, it is necessary to design an intelligent human-computer interaction method which is independent of external devices. The attitude information of human body is obvious and affable to human being. Through computer vision and artificial intelligence technology, the human posture information can be recognized, and the human-computer interaction task can be accomplished. First of all, aiming at the situation that the mechanical vibration of UAV flying in the air causes the shooting picture jitter, the center region template matching method is used to compensate the UAV mechanical vibration in the algorithm. Make the picture more stable and easy to follow up algorithm processing. Then, the human body target in the image is detected. Two solutions are considered, one for earth station and the other for machine processing. The first is the use of convolutional neural network detector (Faster R-CNN), which has complex structure and large computation, to generate candidate regions, classify target types and locate human body targets. This method has high accuracy and can adapt to complex background environment. However, it is difficult to implement real-time on the airborne embedded computing platform. The other is to use the method of extracting foreground information to generate candidate boxes, and then to use convolution neural network with simple structure to complete the classification task. This method needs to model the scene information first, but the computation is small. Real-time performance can be realized on the airborne embedded platform. Then, the attitude information of the detected person is recognized. In order to determine whether the detected person is the object of human-computer interaction, a wave action detector is designed as the switch of the human attitude information detector, which enhances the security of the system. For the objects that have been detected by waving, four kinds of postures are designed, and the UAV realizes the human-computer interaction task by recognizing the types of gestures. In order to improve the running speed and make full use of the processed information, the target foreground information selected by the frame is directly input as a feature into the multi-layer fully connected neural network for recognition. Finally, the algorithm is transplanted to the airborne embedded equipment and flight experiment is carried out. The results show that the system can achieve real-time processing while maintaining high recognition accuracy, and can accomplish human-computer interaction task well.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP11;V279

【参考文献】