Multi-view perception is a fundamental technology for achieving environmental understanding and object detection. With the advancement of artificial intelligence, it has been extensively employed in various applications, including autonomous driving and industrial robotics. However, as an intelligent system, multi-view perception may produce erroneous predictions under certain conditions, potentially resulting in critical failures or accidents. Therefore, it is imperative to perform comprehensive testing of multi-view perception systems to ensure their safety. However, most existing testing approaches predominantly target single-view inputs and lack test cases that account for cross-view consistency. To address this issue, we propose MVTest, an automated metamorphic testing method. It leverages semantic information and spatial constraints to select and optimize object placement. MVTest ensures inserted objects maintain cross-view consistency and their orientation aligns with surrounding road directions. We conducted extensive experiments on four mainstream multi-view perception systems. The experimental results show that: (1) MVTest reveals 32.5% more errors than existing methods, thereby exposing deficiencies in the perception systems’ feature aggregation capabilities; (2) MVTest generates images with a high level of visual realism and scene plausibility, closely resembling the real world.