Year: 2026
Type: article
Abstract: This paper presents DSC-Mamba, a lightweight model for abnormal driving behavior detection from in-vehicle camera images. The model reduces computational load through depthwise separable convolution and designs two key interaction mechanisms: in the DAEDC module, channel attention generates weights to modulate multi-scale spatial attention; in the HEMamba module, the global modeling output of Mamba is normalized and re-injected into local tokens, allowing global context to calibrate local representations. On the SFD3 dataset, using driver-level five-fold cross-validation as the primary protocol, DSC-Mamba achieved an average accuracy of 93.12% ± 0.19% with only 0.421 M parameters and 0.1053 G FLOPs. This is 0.08 percentage points higher than the competitive MVim baseline under the same partition. Under random image-level partitioning, the accuracy reached 99.89%, though this is noted as overly optimistic and for reference only. On the AUCv2 dataset, the model achieved 92.75% accuracy. Furthermore, PyTorchSim-based NPU simulation estimates an INT8 single-frame latency of 12–18 ms and power consumption of 0.9–1.4 W. These results indicate that DSC-Mamba achieves a favorable balance among parameter count, FLOPs, and accuracy, though real-vehicle hardware verification remains future work.
Source: Scientific Reports
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,690
Awards: G453027779, G8180075459
OpenAlex ID: W7234856919
