A method based neural network for pedestrians detection in infrared images is presented. The background image is extracted adaptively with 3D median filtering and regions of interest (ROIs) are segmented out with the image difference method. The Fourier descriptors (FD) are employed as the shape feature of ROIs. A BP neural network is designed to recognize ROIs. The classifier is tested with a lot of samples with different types and different shapes. The experimental results show that the classifier has high true positive rate and low false positive rate, which can distinguish whether the ROIs are pedestrians or not effectively and efficiently.