Abstract: To solve the problem of easily losing the targets when tracking multiple targets by the current video target tracking algorithms，an improved tracking-learning-detection (TLD) algorithm is presented for multi-target detection and tracking by taking armored vehicles and aircrafts in videos as the research objects.For the lost targets，the trajectories of typical targets in video are tracked by using the prediction function of Kalman filtering algorithm，and the tracked trajectories are used to compensate for the lost parts of TLD algorithm so as to obtain the complete trajectories of typical targets in videos，which is beneficial to improve the accuracy of video multi-target tracking.For the poorer accuracy of the existing trajectory prediction methods，a video target trajectory prediction algorithm based on social long short term memory (Social-LSTM) network is proposed. The algorithm integrates the contextual environment information and the interaction relationship among multiple target trajectories into Social-LSTM network and predicts the trajectories of the typical targets to be detected. Simulation experimental results show the trajectory prediction algorithm is superior to the traditional LSTM algorithm，hidden Markov model (HMM) algorithm，and Gaussian mixture model (GMM) algorithm，which is helpful to improve the accuracy of trajectory prediction for typical video targets.

Key words: armoredvehicle, aircraft, objecttracking, multi-targetdetection, social-longshorttermmemorynetwork, trajectoryprediction