Abstract: A new method based on Hessian matrix threshold of finding local low-level saliency features is proposed in this study after the standard local invariant feature extraction algorithm SRUF (Speeded Up Robust Features) is analyzed. In this method, the number of saliency feature points can change with the change of Hessian threshold. The saliency feature points will become sparser when Hessian threshold becomes larger. When a certain extreme threshold which is defined as Hessian Threshold Node is reached, the retained discriminative feature points are remarkable stability characteristics, also make up the best sparse saliency features set. This method is applied in the unmanned vehicle environment sensing system to help extract the before going vehicle’s saliency feature and realize object tracking and obstacle detection. Experiment results show that this is an quick and robust method to determine saliency feature points quantitatively and is very suitable for the occasion which has strong demand on real-time.