In order to reduce the limitation in background statistics estimation of unstructured background detector, a small target detection algorithm based on object-oriented analysis was proposed. After segmenting the whole imagery into many fairly homogenous regions using adaptive iterative method, multivariate normality test was applied to choose several optimal object sets which obey the law of normal distribution well. Then, the selected objects would be combined with GLR to perform target detection. This method could make the local background well fit a normal distribution and effectively separate the target signal from background, and meanwhile avoid the contamination effect through the selection of optimal objects. A simulation experiment was conducted on real OMIS data to validate the effectiveness of the proposed algorithm. The detection results were compared with those detected by the unstructured background detector GLR and improved GLR which incorporated K-Means clustering. The results show that the proposed algorithm has better detection performance and lower false alarm probability than other detection algorithms.