Abstract: ‍ ‍The lidar-based tracking and surveillance of small birds is a new key technology for bird monitoring and providing early warnings at an airport and its surrounding airspace. To address the problems of large target state tracking errors and low model adaptability caused by the low sampling frequency of lidar， this paper proposes an extended Kalman tracking method for target state estimation based on gated recurrent units. By fusing the predictive capability of deep-learning networks for non-linear motion and the noise reduction capability of extended Kalman filtering， this method can track bird targets with non-linear motion， which cannot be accurately modeled under low sampling rate conditions. To address the problem that the deep learning network is an implicit expression model that is difficult to fuse with extended Kalman filtering， an approximate one-step transfer matrix estimation method is proposed to transform the prediction of the deep-learning network into an explicit state transfer model， which makes the prediction and filter estimation of the tracking method form a circular iteration. Simulations on a publicly available pigeon flight trajectory dataset showed that the proposed method outperformed traditional tracking algorithms at different sampling frequencies and provided a tracking error performance improvement of more than 25.5% at low sampling frequencies compared to existing methods.