Vision-based mobile people counting system

People detection and counting systems are highly valuable in multiple situations including managing emergency situations and efficiently allocating resources. However, most people counting systems are based on fixed sensors or fixed cameras, which lack flexibility and convenience. In this paper, we have developed a vision-based mobile people counting system which uses Android smartphones to capture images, and state-of-the-art person detectors, based on artificial intelligence, to count the number of people in a designated area. The embedded devices in smartphones such as camera, clock, GPS, are utilized to provide additional information for data collection. Several person detection frameworks such as You Only Look Once v2 (YOLO2), Aggregate Channel Features (ACF) and Multi-Task cascade Convolutional Neural Network (MTCNN) were evaluated to determine the best performing algorithm capable of offering accurate counting results across different scenarios. The experiments results show that YOLO2 outperforms ACF and MTCNN detection algorithms in different scenarios. However, YOLO2 has its own limitations as it often outputs redundant detections, requiring an additional Non-Maxima Suppression (NMS) algorithm to output a single bounding box per detection. The NMS threshold has to be carefully pre-fixed to provide top detection and counting performance across different scenarios.