This paper presents a vision-based fall detection system to automatically monitor and detect people’s fall accidents, particularly those of elderly people or patients. For video analysis, the system should be able to extract both spatial and temporal features so that the model captures appearance and motion information simultaneously. Our approach is based on 3-dimensional convolutional neural networks, which can learn spatiotemporal features. In addition, we adopts a thermal camera in order to handle several issues regarding usability, day and night surveillance and privacy concerns. We design a pan-tilt camera with two actuators to extend the range of view. Performance is evaluated on our thermal dataset: TCL Fall Detection Dataset. The proposed model achieves 90.2% average clip accuracy which is better than other approaches.

Abstract
 1. 서 론
 2. 선행 연구
  2.1 비영상 기반
  2.2 영상 기반
  2.3 인체 행동 인식
 3. 3D 컨볼루션 신경망
  3.1 3D 컨볼루션 및 풀링 연산
  3.2 신경망 구조 및 학습
 4. 실험 및 결과
  4.1 카메라 팬틸트 제어
  4.2 TCL Fall Detection Dataset
  4.3 실험 결과 및 분석
 5. 결 론
 References

• 김대언(Ujin Technology, Daejeon, Korea) | Dae-Eon Kim
• 전봉규(Robotics Program, KAIST(Korea Advanced Institute of Science and Technology, Daejeon, Korea) | BongKyu Jeon
• 권동수(Department of Mechanical Engineering, KAIST(Korea Advanced Institute of Science and Technology, Daejeon, Korea) | Dong-Soo Kwon Corresponding author