Adaptive Sensor Fusion for Autonomous Driving: A DGConv-Based Approach

Abstract: Ensuring the reliable and safe operation of autonomous vehicles requires advanced perception and real-time data integration from sensors such as LiDAR, radar, and cameras. Each sensor captures distinct, complementary information about the environment, which, when combined, enhances capabilities in object detection, localization, and situational awareness. Traditional multi-stream fusion architectures, though effective, often demand high computational power, potentially compromising real-time responsiveness in dynamic conditions. In this paper, we introduce a single-stream convolutional neural network (CNN) architecture that leverages dynamic grouping convolution (DGConv) to perform efficient and adaptive sensor fusion in autonomous systems. DGConv facilitates the model’s ability to automatically learn optimal configurations for feature-sharing across various sensor inputs, dynamically adjusting fusion depth and structure based on the data properties. This adaptability enables the model to handle a range of sensor modalities and environmental scenarios while striking a balance between accuracy and computational efficiency. Preliminary results show that this DGConv-based single-stream architecture provides a computationally lighter alternative to traditional multi-stream models, maintaining strong multi-sensor fusion performance and improved processing efficiency. By achieving both high accuracy and reduced computational load, this approach offers a scalable pathway for sensor fusion, paving the way for more responsive, reliable, and resource-efficient autonomous driving applications that enhance real-world safety and functionality.