In order to improve the path tracking accuracy of small dual-motor driven crawler mower in orchard under severe road conditions, a path tracking control algorithm based on virtual radar path perception and two-level deep neural network was proposed. Firstly, a two-level serial artificial deep neural network was built, and the first-level deep neural network calculated the relative position relationship between the crawler mower and the target path through the virtual radar path sensing algorithm. The control speed of driving motors on both sides was calculated according to tracking deviation, heading angle, influence factor of lateral deviation, factor of converted track slip rate and relative position relationship between crawler mower and target path, and path tracking control was realized by second-level deep neural network. The U-shaped path tracking tests of crawler mower were carried out on orchard road surface after irrigation. When the vehicle speeds were 0.4m/s and 0.8m/s, the maximum lateral deviations of path tracking algorithm were 0.064m and 0.072m, and the average lateral deviations were 0.026m and 0.033m, respectively. Compared with the traditional pure tracking control algorithm, the maximum lateral deviations at the test speed of 0.4m/s and 0.8m/s were reduced by 31.18% and 20.88%, respectively, and the average lateral deviations were reduced by 35.00% and 29.79%, respectively. The path tracking control algorithm combining virtual radar and two-level deep neural network can effectively improve the track tracking accuracy of crawler mower on bad road surface.