The automatic extraction of the targets which we are interested from a given image is the fundamental of the automatic detection and identification for security screening systems based on imaging technologies. Suffering from the comparatively low signal to noise ratio (SNR), the automatic detection of targets in a passive Terahertz (THz) imaging system facing great challenges, but in urgent necessary. In this paper, a comprehensive method for automatic detection of concealed targets in passive THz image by making the best use of the ‘block statistics uniformity’ properties of the passive images is firstly studied. A theoretical model for the ‘featured regions’ decomposition based on the minimization of a ‘fit energy’ functional with respect to a ‘surface function’ is established, to overcome the drawbacks of conventional methods with gradient-based edge operators for their unsuccessful application in low SNR passive images with blurred boundaries. Based on above theoretical basis and taking advantages of the distinguished contrasts of the convergent ‘surface function’ in different ‘featured regions’, an automatic detection algorithm with three steps was further developed to automatically extract the number, the locations and the shapes of all the concealed targets, with the shape of each target derived as the contour point series arranged in clockwise direction. With plenty of experimental results in 0.2THz band, it's found that, the proposed method has high detection accuracy about 95% with quite good real-time performance, even for the single channel proof-of-state system with low SNR. The theorem, algorithm and results in this paper may have important applications in unmanned and intelligent security screening systems without any artificial interventions.