To solve the problem that single feature can't describe the nonlinear moving targets accurately under complex environment, this paper propopsed a adaptive method which is based on multiple features and particle filter.It first extracted contour direction distribution and color distribution from the target area.And then with help of particle filter theory, it integrated these features to track nonlinear moving targets from the video. In the end, CUDA can provide powerful parallel computing capacity so that this proposed method can be accelerated and meet the need of real-time monitoring. The experiment results of this paper indicates that the proposed method has strong robustness and high accurateness for non-linear and non-Gaussian moving targets. Moreover, it can cope with the situations of targets occlusion or temporarily disappering, focal length transformation and background changing.