Abstract: In order to realize the obstacle avoidance of automatic navigation agricultural machinery and solve the problem that the panoramic camera mounted on the top of the agricultural machinery needs to accurately and quickly detect obstacles in real time to obtain the 360° image information around it, an improved YOLO v3-tiny target detection model was proposed, which can realize the detection and identification of pedestrians and other agricultural machinery in the field. In order to improve the detection effect of small targets in panoramic images, the fast detection speed and lightweight network model YOLO v3-tiny was used as the basic framework, and the splicing layer before the second YOLO prediction layer was used as the third prediction layer by fusing the shallow features with the second YOLO prediction layer to increase the detection effect of small targets; in order to further increase the network model’s ability to extract target features, borrowing the idea of residual network, the residual module was introduced on the YOLO v3-tiny backbone network to increase the depth and learning ability of the network, so that it can better improve the detection capabilities of the network. In order to verify the performance of the model, totally 1 100 original data sets of pedestrian and agricultural machinery obstacles in the farmland environment were established, after data amplification, totally 2 200 images data sets were obtained, the data sets were divided into training set, verification and test set according to 8∶1∶1, and the model was trained under the Pytorch 1.8 deep learning framework. After the model was trained, totally 220 images of test set were used to test different models. The test results showed that the farmland obstacle detection model based on improved YOLO v3-tiny had an average accuracy rate and recall rate of 95.5% and 93.7%, respectively, which were 5.6 percentage points and 5.2 percentage points higher than that of the original network model. Single panoramic image detection took 6.3 ms, the average frame rate of video stream detection was 84.2 f/s, and the model memory was 64 MB. The improved model can meet the real-time obstacle detection requirements of agricultural machinery in motion while ensuring high detection accuracy.