The traditional target detection methods suffer from the quality of target and background training samples, attitude of target, visual angle of target and noise, etc. In order to overcome these limits, a novel method of data-driven quadratic correlation filter based on sparse coding was proposed, in which the dictionary of target autocorrelation matrix is built. This model not only detects target with multiple attitudes and visual angles, but also is insensitive to noise and the quality of training samples. This model is independent of the randomness in different backgrounds. The experimental results on pedestrian and vehicle show that the proposed algorithm is effective. The idea of proposed algorithm is a good reference for improving the methods of filtering.