In the 1970s, Korea began to build urban railroads to cope with increased traffic demand from the development of industries. On August 15, 1974, the first Seoul Metro Line 1 was opened, and Busan opened Line 1 on July 19, 1985. Busan Metro is currently operating a total of four lines, and transporting an average of 900,000 people a day. While the number of transportation personnel has exploded in the past due to rain, the number of employees working in history is decreasing due to automation systems in areas such as ticket sales and gate management, it is therefore necessary to introduce an intelligent CCTV system for absolute safety management of railway stations.
In this thesis, I compare and analyze the detection performance of object detection deep learning models to detect abnormal behavior in urban railway stations in real time, and I also propose a model suitable for embedded edge computing through lightening. The proposed model improves performance by applying dilated convolution to the YOLOv5n version.
As a result of comparing and analyzing the performance of the five versions of YOLOv5 through abnormal behavior learning, it was confirmed that the performance for YOLOv5n version with 155 layers and the x version with 322 layers were not large at 0.6%. Applying the dilated convolution to the YOLOv5n version improves the steady-state detection performance by 3% with some. The dilated convolution solution improves performance for sparse images without increasing the computation cost.
In this thesis, I show that real-time abnormal behavior detection is possible by implementing the YOLOv5n version applied with the dilated convolution as on an Android app. The coefficient file learned with the Pythorch platform is lightweight with the Pythorch lightweight program and applied to the Android app to implement a real-time anomaly behavior detection app to confirm the possibility of applying it to edge computing.
In order to improve the accuracy of abnormal behavior detection, we intended to improve the abnormal behavior detection system developed by expanding the learning data and adding abnormal behavior events that may occur within urban railway stations.