In actual production, the number of wheat seedlings plays a key role in estimation of emergence rate, yield prediction, and grain quality. Timely and accurate estimation of number of wheat seedlings is very important for wheat production. Due to the complex growing environment in the field, imaging of wheat seedlings is easily affected by factors such as illumination, occlusion and overlapping, which results in poor performance when existing target object counting methods were directly used for wheat seedling counting. In order to reduce negative impacts of these factors and further improve counting accuracy, an improved wheat seedling counting model was proposed by enhancing local contextual supervision information based on existing target object counting network, P2PNet (Point to point network). Firstly, wheat seedling images were preprocessed, and a private wheat seedling data set was built by using point labeling method. Secondly, a wheat seedling local segmentation branch was introduced to improve the architecture of P2PNet, so as to extract the local contextual supervision information of wheat seedling. Then an element-by-element point multiplication mechanism was designed to fuse global and local contextual supervision information of wheat seedling. Finally, per-pixel weighted focal loss was introduced to construct the overall loss function, and the model was optimized. Experimental results on the self-built dataset showed that the mean absolute error (MAE) and root mean square error (RMSE) of P2P_Seg were 5.86 and 7.68, respectively, which were 0.74 and 1.78 lower than those of P2PNet. Compared with other state-of-the-art counting models, P2P_Seg exhibited better counting performance. In the actual field environment, the application test analysis, error counting and missing counting analysis were conducted. P2P_Seg was more suitable for complex field environments, and it provided a method for automatic wheat seedling counting.