Abstract:

Accurate environmental perception is fundamental to the realization of Automated Valet Parking （AVP） functions. Traditional AVP systems primarily rely on single-vehicle perception； however， with the continuous development of intelligent technologies at parking facilities， collaborative interaction between vehicles and facilities has become an inevitable trend for the implementation of AVP. A V2X （Vehicle-to-Everything） collaborative whole-areas perception method for underground parking lots is proposed， transforming the global perception challenge into a large-scale graph model problem. By integrating sensor data from facility-side lidar and cameras， along with perception data from connected vehicles， the method establishes various edge constraints based on vehicle poses. To enhance perception accuracy， a large-scale graph model method that incorporates lane-level map information is proposed in this paper， which takes parked vehicles as semi-static constraints while integrates lane-level map data for lateral constraints. A sliding window is introduced during the solving process to reduce the scale of the graph model， with final perception results output in map form for vehicle use. Through simulation experiments and field experiments in underground parking lot scenarios with an area of over 2 500 square meters， the results show that this method significantly improves the perception ability in complex parking lot environment and achieves whole-area perception of underground parking lots.

Key words: underground parking lot, V2X vehicle-parking collaboration, whole-area perception, lane-level map, graph model