To meet the demand for the real-time collision-avoidance detection of close obstacles by an unmanned surface vehicle (USV) on the sea surface, this study establishes an obstacle adaptive grid representation method for the USV based on three-dimensional (3D) lidar. According to the distribution of the lidar point cloud of environmental obstacles around the USV, a functional relationship among obstacle density, obstacle expression time, and grid map resolution is established for adaptively determining the moderate map resolution and constructing a grid map. The 3D lidar point cloud data are subjected to the process of dimensionality reduction and projected onto the grid map to reduce the amount of data and improve obstacle detection efficiency. Furthermore, a method validation experiment is conducted using 3D lidar; consequently, the lidar point cloud data of three different obstacle scenarios are obtained. The results show that the desired resolution of the obtained grid map and number of details regarding the obstacle increase with an increasing number of obstacles. Conversely, the desired resolution of the obtained grid map is lower and obstacle representation is faster with fewer obstacles in the environment, and the obstacle adaptive grid representation can be realized. The follow-up research of USV local collision avoidance path-planning can be supported by the established obstacle adaptive grid representation method.