With rapid advancements in automotive technology, autonomous vehicles are expected to operate widely on roads in the near future. This shift toward automation reduces human responsibility for monitoring surroundings, potentially leading to the Molly Problem, as defined by ITU-T, a scenario where no eyewitnesses are present at the scene. This presents a challenge in seeking alternative data sources to support investigations. Equipped with a range of sensors, vehicles functioning as witnesses or “witness vehicles” have gained attention in recent studies for supporting evidence gathering. However, casting a wider net to expand the potential pool and effectively identifying relevant witnesses within it is a challenge. According to existing works, witness vehicles are identified by those within a specific range receiving alert messages at the time of the incident. However, due to the dynamic nature of traffic and delays in accident detection and verification, actual witnesses may have already left the scene by the time these messages are broadcast. Without effective approaches for identifying actual witness vehicles, this could result in the collection of irrelevant data, further complicating the investigation process and posing privacy concerns for the data owners. To address these issues, we propose a “call-for-witness” protocol and four witness identification approaches: Proximity Assessment, Perception Field-Target Overlap Assessment, Visual Object Detection, and LiDAR-Guided Object Detection. The traffic authority can specify search queries, which vehicles use to perform local assessments, allowing them to determine if they possess data relevant to the incident. The implementation and evaluation were conducted on data collected from the CARLA simulator, which was used to model vehicle collisions with varying traffic densities. The results demonstrated significant improvements in reducing the amount of data and increasing the relevance of the data submitted to the traffic authority.