Autonomous driving necessarily involves mapping the surrounding terrain. Various algorithms are used, supplemented by the odometry or GPS signals. In the case of driving through very difficult terrain as woods, cart tracks, where this information is subject to large errors, only Lidar information needs to be used. The article deals with processing the Lidar data from this extremely difficult terrain, which contains very few features suitable for registration. The main result is the implementation and evaluation of the behaviour of algorithms for the reconstruction of the surroundings of a vehicle in this challenging environment. The work presents a comparison of three keystone algorithms for the point cloud registration, i.e. based on the distance of the points (Iterative Closest Point), based on the statistical evaluation of the neighbourhood (Normal Distribution Transform) and based on the description of the important points using a histogram (Fast Point Feature Histogram). We show four different scenarios and the quality of result is computed by error evaluation.