Abstract: With the rapid development of real-life 3D city construction, electronic maps can now be presented in 3D real-life scenes, and use more intelligent image processing methods to integrate the real world and real-world model data to achieve the combination of virtual and reality. In this context, in view of some problems in the detection method of electronic fence in 2D scenes, this paper proposes a cross-border detection method of electronic fence in the case of the fusion of video and three-dimensional real scene. The underlying data is based on the real-world 3D model and the surveillance video stream, firstly, the virtual 3D space electronic fence projected by video surveillance in the geographical scene is established, and at the same time, the video is connected to the optimized deep learning neural network model and the human posture estimation is applied, and the estimated point coordinates and fence coordinates are converted into 3D local coordinates through coordinate conversion, and then the detection algorithm is compared with the planned 3D electronic fence in real time to realize the real-time and effective judgment of the object crossing the boundary. Through experimental verification in different video scenarios, the results show that the method is effective and feasible, without specific hardware support and scene constraints.

Key words: video fusion, electronic fence, deep learning, coordinate conversion, out-of-bounds detection