Aimed at the problems of dynamic target tracking failure in the situation of fast motion and target occlusion, a spatiotemporal context tracking algorithm is proposed based on key points. Firstly, the key points of the target are extracted, and the predicted location of the object is obtained by combining key points matching with optical flow tracking. Then, the relationship model between key points change rate and spatiotemporal context model updating rate is established to control the update rate in realtime. In this way, the introduction of erroneous information can be prevented. Finally, a local context appearance model is constructed in the predicted location region, and the correlation between the spatiotemporal context model and the local context appearance model is computed to obtain the confidence map. Furthermore, the target is located accurately. The algorithm is validated in the test video, the highest average tracking success rate is 60% and the minimum average center error is 26.14 pixel. Compared to the 4 types of current major algorithms, the comprehensive performance of the proposed algorithm is superior to other methods, whose average tracking success rate is 90%, average center point error is 7.47 pixel and the average tracking rate is 25.31 frames per second. In the case of background interference, occlusion, target rotation and rapid motion, the mobile robot with binocular vision is used to track the random moving target. The success rate is 97.4%, and the average relative error of tracking distance is 4.05%.