Abstract: In the fields of computer vision and close-range photogrammetry,marker points are widely used,and their positioning and accurate extraction directly affect the observation accuracy.However,in long-term monitoring,complex lighting conditions can lead to poor recognition and extraction effects of marker points,thereby affecting the monitoring accuracy.For this purpose,this paper proposes a marker point extraction method based on deep learning algorithms.Firstly,a marker point dataset is established by using marker point images in different lighting environments.Then,the marker points are identified under different lighting conditions and accuracy analysis is conducted.Finally,displacement experiments were conducted on the marker point areas extracted by the YOLOv7 algorithm to determine the observation accuracy.The results show that the YOLOv7 deep learning algorithm can quickly and accurately identify the region of interest of the marker points,with a mAP of 95.45%,an F1 value of 94.36%,and a frame rate of only 4.40.Under different lighting conditions,the landmark area can be accurately identified and the observation accuracy is high.The research results can provide effective solutions for the automatic and high-precision extraction of marker point areas in complex environments in long-term dynamic monitoring.

Key words: YOLOv7 algorithm, long time series monitoring, marker point, complex illumination, region of interest extraction