Intelligent robot has recently matured in industry, whose core technology is machine vision, especially object recognition. In existing object recognition methods, scenes are segmented based on color, and features are then extracted to recognize objects. However, over segmentation exists for scenes with complex color features, which influences subsequent object recognition process. To deal with the over segmentation problem, a multi-object image segmentation algorithm based on local features is proposed, which uses binocular camera to collect scene images. Firstly, the scene image is preprocessed. The depth information of the scene is then obtained by stereo matching, and is used to determine the target area. Secondly, the local features of the target region are extracted by a scale-invariant feature transform (SIFT) algorithm with dynamic threshold, and the local features are then transformed into feature constraints. Finally, the feature vectors, which are based on region constraint, feature constraint and spatial information, are used for clustering segmentation to obtain the final segmentation result. Simultaneously, each object region is recognized. The experiment results show that the overall error rate of the proposed algorithm is less than 10% for a scene with complex color features, and is reduced by 15% compared with those of the existing algorithms.