Efficient Collaborative Perception with Adaptive Communication in Bandwidth-Constrained Scenarios

In autonomous driving, collaborative perception significantly improves safety by addressing occlusions and expanding the perceptual range of individual vehicles. Nevertheless, effectively exchanging data among vehicles under bandwidth constraints presents a major challenge, especially regarding adaptive communication strategies. To tackle this, we introduce Adacomm (Adaptive Communication for Collaborative Multi-agent Perception), a network-adaptive framework that enhances multi-agent collaborative perception by optimizing data exchange to improve perception accuracy and reduce communication costs. The Adacomm framework consists of modular operations such as blind spot identification, multi-scale feature generation, data scheduling optimization, shared message fusion, and detection decoding. Experimental results demonstrate that Adacomm enhances multi-vehicle perception by 16.8% in AP@IoU=0.5, compared to traditional single-vehicle perception, and lowers average communication costs by 29.5% relative to the existing Where2comm method, with a minimal performance decline of 5.6%. Our analysis across different feature layers indicates a trade-off between communication volume and perception accuracy, with medium-resolution features providing a balanced solution. Overall, Adacomm effectively addresses critical challenges in collaborative perception, showing substantial benefits and advancements in scenarios with limited communication resources.