T-CLOCs: Transformer-Based Camera-LiDAR Decision-level Fusion for Object Detection of Robot in Outdoor*

Robust object tracking based on multi-sensor data represents a critical technology for robot in outdoor tasks. Current monocular sensor architectures exhibit inherent limitations including insufficient tracking precision, susceptibility to environmental perturbations, pronounced noise contamination artifacts and constrained adaptability in complex operational scenarios. While multi-sensor fusion offers promising solutions, it presents new challenges in achieving efficient fusion algorithm implementation alongside lightweight deployment on resource-constrained embedded systems. This paper proposes a novel decision-level fusion approach employing a hybrid Transformer-CNN architecture. First, we replace the dual convolution structure with a two-layer encoder to enhance accuracy. Second, we introduce a dynamic subsampling strategy to adaptively compress intermediate tensors. Experimental comparisons with the baseline CLOCs method on KITTI dataset indicate that our algorithm attains 98.74% accuracy (2.48% improvement) in AP@0.7 evaluation under Easy conditions, while preserving comparable performance at 90.62% under Moderate conditions. The experimental results validate the effectiveness of the proposed algorithm. The method also demonstrated excellent performance in real-world robotic deployment scenarios.