Obstacles such as large irregular fruit tree canopy and pedestrians are hidden dangers that hinder the movement of mobile picking robots in orchard environments. To improve the safety of mobile picking robot in the orchard environment and preventing the collision between mobile picking robot and obstacles, this study took the spindle-shaped apple orchard as the research object and proposed a mobile robot navigation path optimization method based on the improved artificial potential field. The mobile picking robot was composed of robot chassis, LiDAR (RS-LiDAR-M1, Suteng Innovation Technology Co., Ltd. China), the main controller (Jetson TX2, NVIDIA Co. Ltd., USA), and the picking arm, the main controller installed the ubuntu16.04 LTS operating system, and developed the program based on Robot Operating System (ROS) and Point Cloud Library (PCL). The main steps of the mobile robot navigation path optimization method based on the improved artificial potential field included point cloud preprocessing, extraction of ridgelines, and optimization of initial path. Firstly, the LiDAR carried by the mobile picking robot collected the three-dimensional point cloud between the rows of the orchard. The three-dimensional point cloud was processed through filtering, down-sampling filtering, and statistical filtering, and the ground plane algorithm removed the orchard ground point cloud and extract the orchard ridge and fruit tree canopy point cloud. Secondly, the Least Square Method (LSM), Hough Transform, and Random Sampling Consensus (RANSAC) extracted the ridgelines from the orchard ridge point cloud. The middle lines of the two sides of the ridgelines were used as the initial path. Finally, the artificial potential field was improved by discarding the gravitational potential field and established the potential field of the fruit tree canopy profile point cloud. The initial path was discretized into discrete points, and each discrete point was optimized in turn according to the improved artificial potential field, and then the optimized discrete points were fitted by quadratic B-spline curve to obtain the optimized path which had the ability to avoid large fruit tree canopy and pedestrian obstacles. The results of the extracted initial path by LSM, Hough transform, and RANSAC were analyzed in terms of real-time performance and anti-noise ability. The results showed that the three methods successfully extracted ridgeline and initial path, and RANSAC had the best real-time performance and the average processing time was about 0.147 × 10-3 s, the standard deviation was 0.014×10-3 s, and it had good anti-noise ability. Based on the extracted initial path by RANSAC, the improved artificial potential field method was used to optimize the initial path, which avoided the problem that the traditional artificial potential field method was easy to fall into oscillation. The shortest distance between the obstacle point cloud and the navigation path was increased from 0.156 m to 0.863 m, and the average optimization time and the standard deviation was 0.059 and 0.007 s, respectively, which indicated the optimization method basically had the ability to optimize the path in real-time to avoid obstacles. The method of optimizing the navigation path proposed in this study basically could meet the requirements of safety and real-time, which could provide a technical reference for the autonomous navigation of mobile picking robots in the orchard environment.