室内场景下服务机器人的移动作业，对于行人的检测算法提出了更高的要求。不同于室外场景，室内场景存在遮挡、尺度、行人姿态复杂等问题，同时移动机器人的感知任务不仅要检测出行人的位置，还要检测出行人具体的属性，如年龄、朝向等，以便更好地服务。此外，部署在移动机器人上的算法对于模型的推理速度有着极高的要求，在较快运算速度的同时要保证模型的检测效果。因此，本硕士论文提出了室内场景轻量化行人属性检测方法，并初步应用到室内移动平台上。本文的工作与贡献如下：首先，本文构建了室内场景行人属性数据集（IPA数据集）。在移动平台下使用视觉传感器进行数据采集，覆盖商场、服务大厅、餐厅等场景，进而进行数据的清洗、标注与整理，针对类别不平衡问题进行了数据的增强。之后从数据特性与训练过程对数据集可行性进行评估，以证明数据集利于训练。最后在IPA数据集上进行初步的行人属性检测实验，当下主流的方法出现了误检漏检、过拟合等问题。其次，本文针对室内场景行人属性检测的实际效果不佳问题，设计了轻量化行人属性检测网络。重点设计了注意力机制融合、轻量化损失与迁移训练的方法，以提高模型的检测效果，同时保证模型推理速度。在IPA数据集上，本文的方法比现有主流方法的总类别平均精度提高了6.2%，同时检测速度提高了84%。在实际场景的测试试验中，模型整体效果较好，对于近距离目标零误检。最后，本文将行人属性检测方法初步应用到仿真系统与移动机器人平台上。首先建立行人属性检测系统框架，随后搭建了基于Unity的室内场景，设定行人姿态与运动状态，以及相机视角，进而部署算法到该系统中，测试算法的性能。最后在移动机器人平台上，测试算法在实际场景中的表现效果，众多结果表明模型的效果与速度满足服务机器人的作业要求。
The mobile operation of service robots in indoor scenes puts forward higher requirements for pedestrian detection algorithms. Different from outdoor scenes, indoor scenes have problems such as occlusion, scale, and different postures of pedestrians. At the same time, the perception task of mobile robots must not only detect the position of pedestrians, but also detect specific attributes of pedestrians, such as age, orientation, etc., in order to better ground service. In addition, the algorithm deployed on the mobile robot has extremely high requirements for the reasoning speed of the model, and the detection effect of the model must be guaranteed while reducing the number of calculations. Therefore, this master‘s thesis proposes a lightweight pedestrian attribute detection method in indoor scenes, and initially applies it to indoor mobile platforms. The work and contributions of this paper are as follows:First, this paper constructs an indoor scene pedestrian attribute dataset (IPA dataset). Under the mobile platform, visual sensors are used for data collection, covering shopping malls, service halls, restaurants and other scenes, and then data cleaning, labeling and sorting are carried out, and data enhancement is carried out for the problem of category imbalance. Afterwards, the feasibility of the data set is evaluated from the characteristics of the data and the training process to prove that the data set is conducive to training. Finally, a preliminary pedestrian attribute detection experiment was carried out on the IPA dataset. The current mainstream method has problems such as false detection, missed detection, and overfitting.Secondly, this paper designs a lightweight pedestrian attribute detection network for the poor practical effect of pedestrian attribute detection in indoor scenes. The focus is on the design of ECA attention mechanism fusion, lightweight loss and migration training methods to improve the detection effect of the model while ensuring the speed of model reasoning. On the IPA dataset, the method in this paper improves the overall class average accuracy by 5.2% and the detection speed by 38% compared with the existing mainstream methods. In the test experiment of the actual scene, the model has zero false detection for the close-range target.Finally, this paper initially applies the pedestrian attribute detection method to the simulation system and the mobile robot platform. Build an indoor scene based on Unity, set the posture and motion state of pedestrians, and the camera perspective of the mobile robot, deploy the algorithm to the system, and test the performance of the algorithm. Then, on the actual mobile robot platform, test the performance of the algorithm in the actual scene, and many results show that the effect and speed of the model meet the operating requirements of the service robot.
