In the process of obstacle avoidance control and path planning of autonomous agricultural vehicles, it is not enough to only detect obstacles. If the obstacles are classified and identified more accurately with types and danger degree, it can be served as an important basis for controlling the speed of agricultural vehicles and planning the trajectory during obstacle avoidance, and it will make the process of obstacle avoidance control and path planning of autonomous agricultural vehicles more accurate and reasonable. A convolutional neural network model was designed to realize the identification of multiple obstacles for unmanned agricultural vehicles. The model included an input level, two convolution levels, two pooling levels, a full connection level and an output level. And databases of human and agricultural vehicles were built according to MIT database and website data respectively. Then, the convolution operation was performed on the training set with 5×5 convolution kernel and the acquired feature graphs were pooled in a 2×2 neighborhood. After the convolution operation with 3×3 convolution kernel and 2×2 neighborhood pooling operation, the network model parameters were obtained by using automatic learning and the optimal network identification model was achieved. The experimental results showed that high recognition accuracy of obstacle identification was obtained, which was 94.2%.