In this paper

we propose a new method for foreground object detection based on the Kernel: Density Estimation of a local spatio-temporal model (LST-KDE)

which overcomes information redundancy and the large calculated quantity problem in the training phase as well as the manual adjusting time window size and shadow problem in the detection and updating background phase. The LST-KDE algorithm uses the k-means clustering algorithm to optimize the sample set and to choose the key frames in the training phase. Therefore

it can avoid information redundancy and the large calculated quantity problem. In the detection and updating background phase

the LST-KDE algorithm constructs a local spatio-temporal model. This method can not only adaptively set the time window size by using history frame information in a temporal model

but also uses color and texture features described with the local binary pattern (LBP) algorithm to remove shadows in the spatial model. The experiment in a complex environment demonstrates that the proposed method outperforms recent state-of-the-art methods.