Speed bump detection through computer vision and deep learning is essential for
advancing active suspension preview control and intelligent driving. Although
substantial progress has been made in this field, there remains a need to
enhance detection accuracy while reducing computational demands. This article
introduces a novel single-stage speed bump detector, the Speed Bump Detector
Based on You Only Look Once (SBD-YOLO), which utilizes the YOLOv9 architecture
for speed bump identification. To better capture the deep global features of
speed bumps, we propose an innovative convolutional module—specifically, a
lightweight building block designed for efficient feature extraction—named the
Aggregated-MBConv. Furthermore, we design a new YOLO backbone by stacking Mobile
Inverted Bottleneck Convolution (MBConv) and Aggregated-MBConv modules, which
reduces computational cost while enhancing detection accuracy. Additionally, we
introduce a Squeeze-aggregated Excitation (SaE) attention mechanism at the
network’s neck, which, through parallel operation, enables collective
integration across branches, further improving network performance. A dedicated
speed bump dataset was created to validate SBD-YOLO’s effectiveness. Compared to
YOLOv9, SBD-YOLO achieves a 9.3% increase in precision, a 2.5% boost in recall,
and improvements of 2.2% and 1.4% in mean Average Precision at an
Intersection-over-Union (IoU) threshold of 50% (mAP50) and mean Average
Precision over IoU thresholds from 50% to 95% (mAP50-95), respectively.
Moreover, the number of parameters is reduced by 5 million, and computational
complexity is decreased by approximately 82.8%. These results demonstrate the
significant potential of SBD-YOLO for active suspension preview control.
