基于改进的单次多盒检测器室内人数检测

Indoor People Counting with Improved Single Shot Multibox Detector

室内人数检测是解决公共资源合理分配和利用问题的关键。针对室内人群分布复杂且存在相互遮挡,而传统图像处理算法的准确率较低的问题,使用单次多盒检测器(single shot multibox detector, SSD)结合MobileNetV2与SENet的深度学习目标检测方法,对室内环境下的人进行识别。在微软开源数据集(common object in context, COCO)的基础上,采集室内真实图像制作数据集,进行不同IOU阈值、不同拍摄角度条件下的实验,并部署到计算环境为搭载神经元计算棒(neural compute stick,NCS2)的树莓派。实验表明,改进SSD目标检测模型在IOU阈值为0.4下,平均准确率和召回率较高,分别为97.91%和90.72%,在此计算环境下检测速度可达8帧/s,模型具有良好准确率和实时性。

Indoor people counting is the key to solving the problem of rational allocation and utilization of public resources. For indoor situations, various crowd distribution and mutual occlusion resulted in low recognition accuracy in traditional image processing algorithms. To solve the above problems, an indoor population statistics method was proposed based on single shot multibox detector(SSD), in which the MobileNetV2 and SENet were combined as the basic feature extraction. A dataset was created based on the Microsoft Open Source in Context(COCO) and indoor real images collected. Experiments with different IOU(intersection-over-union) thresholds at different camera angles were conducted, and the computing environment as a Raspberry Pi with a Neural Compute Stick(NCS2) was deployed. Result show that the improved SSD object detection model had a higher average accuracy and recall rate of 97.91% and 90.72%, respectively, at the IOU threshold of 0.4, and the recognition speed could reach 8 frames per second in this computing environment. The model has a proven good accuracy and real-time performance.