This paper focuses on multi-objects tracking problems in monocular camera and mutative background complex scenes. We achieve multi-objects tracking in dynamical scenes under a semi-supervised learning framework, which combines tracking-by-detection method and particle filter together, integrates unreliable information sources and extracts a high-confidence observation model from it. Then an objects maintenance scheme is carried out to cope with occlusion, background changing, entry/exit and so on. The results on standard datasets demonstrate advantages of the proposed algorithm in complex environment, particularly on scenes with occlusion and obstruction.