Aiming at the key point detection problem of UAV, a key point detection algorithm based on cascade neural network is proposed. The neural network used in the algorithm is cascaded into two parts: network1 detects the whole target; network2 receives the target image as input, then outputs the location information of the key points on the target. Aiming at the low real-time problem caused by the existing methods by deepening the accuracy of the network, this algorithm introduces two kinds of cross-level connection methods to enhance the reuse of global information and improve the accuracy of key point positioning. At the same time, it reduces the amount of network parameters and improves the real-time performance by using depth separable convolutions. The test data shows that the relative error of key point positioning is 0.03 in complex background, and the average running speed is 28 f/s on Nvidia Geforce GTX 1080ti. While guaranteeing high positioning accuracy, it meets the real-time requirements of current applications.