一种基于YOLOv8的轻量化盲区检测网络

A lightweight blind spot detection network based on YOLOv8

近年来,全国的交通安全形势日益严峻,交通事故频繁发生,人员伤亡和财产损失惨重。其中,因视觉盲区受限引起的人车碰撞事故最为常见,由于传感器的高昂造价和在盲区检测方面的研究应用较少,预防此类事故主要依靠司机驾驶经验。针对盲区检测和研究的不足,提出一种简洁高效的轻量化盲区检测网络BsDet和BsDet+。轻量化网络以最先进的YOLOv8为基础,结合其他YOLO网络的优点,在头部和颈部进行了轻量化重构,在特征提取部分使用改进的深度可分离卷积降低网络的参数量与计算量。在特定层使用更大的卷积核来扩大感受野,进一步提高网络的检测精度。在构建的盲区数据集进行实验,实验结果表明,BsDet拥有97.72%的mAP和300.76 f/s的FPS,BsDet+的mAP和FPS分别为99.35%和181.31 f/s,相比于SOTA方法,提高了36.8%的检测速度和1.44%的mAP。两种网络分别在树莓派、安卓手机和便携式计算机上进行部署与测试,结果显示在任何平台上,BSDet均拥有最高的检测速度。BsDet和BsDet+可适用于不同性能的硬件与检测需求,具有设备要求低、准确率高、速度快等特点,不仅为轻量化设计提供了借鉴,也能够有效改善基于视觉的辅助驾驶技术。

In recent years,the national traffic safety situation has become increasingly severe,with frequent traffic accidents,heavy casualties and property losses.Human vehicle collision accidents caused by visual blind spot limitations are the most common.Due to the high cost of sensors and limited research applications in blind spot detection,prevention of such accidents mainly relies on the driver's driving experience.In allusion to the shortcomings of blind spot detection and research,a concise and efficient lightweight blind spot detection network BsDet and BsDet+is proposed.The lightweight network is based on the most advanced YOLOv8 and combined with the advantages of other YOLO networks to perform lightweight reconstruction on the head and neck.In the feature extraction section,an improved depthwise separable convolution is used to reduce the parameter and computational complexity of the network.The larger convolutional kernels at specific layers are used to expand the receptive field and further improve the detection accuracy of the network.The experiments were conducted on the constructed blind spot dataset,and the results show that BsDet has 97.72%mAP and 300.76 f/s FPS,while BsDet+has 99.35%and 181.31 f/s mAP and FPS,respectively.In comparison with the SOTA method,it can improve detection speed by 36.8%and mAP by 1.44%.Two types of networks were deployed and tested on Raspberry Pi,Android phones,and portable computers,and the results show that BSDet has the highest detection speed on any platform.BsDet and BsDet+can be applied to different hardware and detection requirements with low equipment requirements,high accuracy,and fast speed.They not only provide reference for lightweight design,but also effectively improve visual assisted driving technology.