In order to detect multi-motion targets in time sequenced forward looking infared (FLIR) images, a novel automatic detection method is proposed. First, the speed of 3D spatiotemporal texture vectors is used to measure motion features. Then, the saliency of motion target is defined under the frame of visual attention. So salient moving objects are detected and extracted by accumulating directionally-consistent optical flow. The regions are regarded as final visual attention regions. An improved geodesic active contour model is built, where the topology-free deformable evolving curve is attracted by the boundaries of attention regions. A gradient based convergence term forces the curves evolution to stop on the object boundaries. The validity of the proposed method is proved experimentally and the results of multi-dataset are given. Compared to other traditional methods, the proposed method has a good noise tolerance and can handle fast or slow image variations. It has no response to unnecessary motion and only focuses on the salient motion targets. The multiple targets boundaries are located due to implementation of level set scheme, which gives a reliable input for targets identifying and tracking or 3D reconstruction and representation.