Real-Time Target Detection in Compressed Domain for Streak Tube LiDAR by Two-Pass Labeling and Sparse Attention

Airborne streak tube imaging LiDAR (ASTIL) enables high-frame-rate 3D imaging but suffers from real-time processing bottlenecks due to massive data throughput and costly full decompression. We propose a compressed-domain detection framework that directly processes natively group-sparse (GS) encoded streak images. The pipeline integrates three components: (1) flag-grid-guided selective decoding for zero-overhead signal extraction; (2) an O(N) Two-Pass Connected Component Labeling (CCL) algorithm replacing DBSCAN; and (3) a 6724-parameter Sparse Set Attention Network (SSAN). Crucially, the SSAN synergizes a 22-D physically grounded feature vector with a dual-prototype cross-attention mechanism, implicitly decoding the bimodal scattering signatures of ASTIL targets. Evaluated on 17,528 airborne frames, our method achieves a Pareto-optimal trade-off, attaining 73.3% Grouped F1-score (G-F1), a diagnostic metric that merges same-class fragments before matching, at 1536 FPS on a single CPU core. This represents a transformative speedup—≈230× faster than PointNet++ and ≈120× faster than Faster R-CNN—while maintaining competitive fidelity. Ultimately, this work demonstrates that aligning algorithmic design with the intrinsic physics of sparse modalities effectively bridges the accuracy-throughput chasm for next-generation real-time airborne remote sensing.