| 摘 要: 交通安全监控图像中的汽车驾驶员安全带检测,可以协助交通管理部门规范驾驶员的驾驶行为。针对汽车驾驶员安全带检测中目标尺寸较小、特征对齐难度较大、检测速度较慢等问题,在MobilenetV2轻量化主干网络的基础上,引入Oriented RCNN 旋转框目标检测算法,提出注意力特征融合模块(Attention Feature FusionModule,AFFM),从而构建了一种轻量高效的端到端旋转框安全带检测算法。该算法的平均精度(AP)达到0.905,查全率(Recall)达到0.949,参数量(Params)仅需要18.54 MB,端到端的检测推理速度(FPS)达到每秒14.6张图片。实验结果表明,该算法有效提高了监控图像中汽车驾驶员安全带检测性能,在实际应用中具备一定的竞争力。 |
| 关键词: 安全带检测;旋转目标检测;注意力机制 |
|
中图分类号: TP391.41
文献标识码: A
|
|
| Research on Lightweight and Efficient Seat Belt Detection Algorithm Based on Rotating Frame |
|
SONG Changming, LIANG Chaoyang, XIAO Lu, SONG Meng, CAI Shuo
|
(College of Science, Zhongyuan University of Technology, Zhengzhou 450007, China)
cmsongh@163.com; 2020003053@zut.edu.cn; 2020103058@zut.edu.cn; songmeng3057@163.com; 1049636164@qq.com
|
| Abstract: Driver's seat belt detection in the traffic safety monitoring image can help the traffic management department to regulate the driver 's driving behavior. Aiming at the problems of small target size, difficult feature alignment and slow detection speed in the detection of automobile driver's safety belt, this paper proposes a lightweight and efficient end-to-end rotating frame safety belt detection algorithm, by introducing the Oriented RCNN rotating frame target detection algorithm and proposing the Attention Feature Fusion Module ( AFFM), based on the MobilenetV2 lightweight backbone network. The average precision (AP) of the algorithm reaches 0.905, the recall rate (Recall) reaches 0.949, the parameter quantity (Params) only requires 18.54 MB, and the end-to-end detection and inference speed (FPS) reaches 14.6 images per second. The experimental results show that this algorithm effectively improves the detection performance of car driver safety belts in monitoring images, and has certain competitiveness in practical applications. |
| Keywords: safety belt detection; rotated object detection; attention mechanism |
