Adaptive optimization strategy and evaluation of vehicle-road collaborative perception algorithm in real-time settings

Abstract

The Intelligent and Connected Vehicle Cloud Control System is a critical approach for achieving high-level autonomous driving. One of the key challenges at the perception level is utilizing multi-source sensory data to create a real-time digital twin of the transportation system. Collaborative perception technology plays a pivotal role in addressing this challenge. However, most prior research has been conducted offline, where the focus has primarily been on comparing ground truth at the sensing timestamp with the algorithm’s predicted perception values. This approach tends to prioritize computational accuracy, neglecting the fact that the physical world continues to evolve during the processing time, which can result in an accuracy drop. As a result, there is a growing consensus that both latency and accuracy must be considered simultaneously for real-time applications, such as digital twins and beyond. To address this gap, we first analyze the comprehensive time delay problem in vehicle-road collaborative perception algorithms and formally define the real-time perception problem within this context. Next, we propose an adaptive optimization strategy for vehicle-road collaborative perception, which accounts for the complexity of the perception environment and the vehicle-road communication pipeline. Our approach dynamically selects the optimal model parameter set based on the perception scenario and real-time communication conditions. Experimental results demonstrate that our strategy enhances real-time performance by 5.8% compared to the best global single-model algorithm and by up to 27.5% compared to the conservative fixed single-model approach.