This paper presents a parking lot obstacle detection system that utilizes infrastructure-based cameras to detect and track obstacles within a parking lot. The system begins by capturing images from ceiling-mounted cameras and employs object detection algorithms to identify 2D detection boxes and estimate 3D bounding boxes’ ground points. By leveraging camera intrinsics and extrinsics, along with the assumption that all objects are on a plane, we transform these points into the world coordinate system. The system then fuses detection results from multiple cameras using spatial and temporal fusion techniques, including the mean shift algorithm and the Expanded Kalman Filter, to optimize and refine the detection results. Our approach effectively addresses the challenge of depth estimation in monocular camera setups and provides a robust solution for parking lot obstacle detection.