Lightweight framework for the violence and falling-down event occurrence detection for surveillance videos

Hyungmin Kim, Hobeom Jeon, Dohyung Kim, Jaehong Kim

Abstract: Recently, there are huge advances in the computer vision and deep-learning technology. However, the practical and light-weight solution for capturing the abnormal event is still less researched, especially in bad weather conditions such as snow and rain. In this paper, we propose the entire pipeline of the abnormal event occurrence detection system for realistic surveillance videos. The proposed system can detect two abnormal behavior: violence and falling down. While modern object detection models accomplish remarkable performance, pedestrian detection and tracking with fast processing speed still have problems, primarily when abnormal behaviors such as falling down occur. In this paper several tracking enhancement method is proposed. The proposed methods improve the F1 score of 21.21% of the detection of the falling down on the largescale CCTV dataset named KISA overseas. Finally, the proposed system achieved the reliable performance of 91% averaged F1 score on the KISA-v2 test set which address the various weather condition and places.