摘要：Object detection has always been a research topicin computer vision. With the deep learning, conventional objectdetection can obtain satisfactory results. However, some objectdetection problems with serious occlusions still need to beimproved. This paper proposes a high accuracy detection andalarm method based on an improved YOLOv3 network andmulti-frame direction field and solve the serious occlusionsproblem. Meanwhile, we apply the method into the elevatorscene to solve the elevator security. Firstly, an improvedYOLOv3 network is proposed for enhancing the featureextraction ability of the network and increasing the recognitionaccuracy. Then, multi-frame direction field is proposed todecrease false alarm. The experimental results havedemonstrated that our proposed method can effectively solveobject detection problem with serious occlusions, and has muchpracticality value in elevator scenes and other scenes.
关键词：E-bike intrusion; YOLOv3; Multi-frame direction field; Detection and alarm
