In space-based optical systems, during the pixel-plane tracking for closely spaced objects (CSOs), in traditional methods, pixels are partitioned after constant false alarm rate detection (CFAR), where higher false alarm rate results in more clutter measurements while lower false alarm rate results in the loss of targets’ information. To solve this problem, CSOs’ feature on pixel-plane were analyzed and a pre-detecting method using Markov random field model(MRF) was proposed. Then pixels were partitioned with k-means. Simulations indicated that detection and partition with MRF provides higher performance than traditional method, especially when signal-noise ratio is poor.