The safety envelope, defined as the theoretical boundary of passenger body parts around the seat during operation, represents a critical design consideration in avoiding collisions between passengers and external objects, and is a fundamental aspect of roller coaster safety engineering. To address the challenge that the intrusions of the safety envelope cannot be adequately identified through unsupervised methods, this study proposes a cluster screening method for point cloud data. The multi-frame point cloud data obtained from a single operation are fused to form a background heat map, which is then superimposed with the objects in order to calculate their heat value. Subsequently, the objects with a lower heat value are labeled as intrusions. The method was validated through an experiment in which a data set obtained from a real roller coaster using a LiDAR detection system was processed. The experimental results demonstrate that the novel method is capable of accurately distinguishing between intrusions and background stationary objects, such as chassis and seats, in an unsupervised setting. Additionally, due to the nature of unsupervised learning, the new method facilitates the implementation of the intrusions recognition function on arbitrary coasters without the need for manual labeling. The proposed method could be utilized for the monitoring of collision protection conditions associated with the operation of roller coasters.