Overhead view based person counting using deep learning

Kaw, Chee Zhao (2022) Overhead view based person counting using deep learning. Final Year Project, UTAR.

Preview

PDF Download (1564Kb) | Preview

Abstract

Detecting people in an image or a video has become more prevalent due to the rapid advancement of technologies in the field of artificial intelligence. In conventional video surveillance systems, most of the person detection methods are based on frontal view, which may have lower accuracy stemming from the occlusion problem. This project proposes an overhead view based person counting system by enabling wider scene coverage and visibility. The entire project methodology can be divided into several phases. First, the YOLOv4 and YOLOv4-tiny object detection models are trained with the dataset of overhead camera perspective. Second, the OpenVINO Inference Engine is utilized to optimize the trained models in order to facilitate real-time implementation. Third, the accurate tracking of each detected person is performed using the deep learning based tracking framework, known as DeepSORT. Lastly, the performance of the proposed system is benchmarked based on the detection accuracy, frames per second (FPS) and counting accuracy. Based on the results obtained, the YOLOv4- tiny model is chosen as it can achieve high fps without the need of high processing power. Besides, the Centroid Tracking algorithm achieves around 38.4% to 40.4% higher fps as compared to that of the DeepSORT tracking algorithm. However, the counting accuracy of Centroid Tracking algorithm is about 22.2% lower than the DeepSORT tracking algorithm. Hence, the overall performance of the YOLOv4-tiny model integrated with DeepSORT algorithm outperforms the other tracking algorithms.

Actions (login required)