目前电动车头盔佩戴监督主要依靠人工，这会造成人力成本的浪费。我国电动车安全驾驶事故中约80%为颅脑损伤致死，随着智能安防的快速发展和随着国家提出的“一盔一带”政策的实施，目标检测技术在电动车头盔领域得到应用。本文通过对一些经典目标检测算法进行分析，选择SSD(Single Shot MultiBox Detector)作为本文研究的基础算法。

针对电动车头盔检测问题，首先在SSD基础上设计了一种可实现二次检测的级联网络CSSD。将人与电动车整体作为一次检测，再对检测后的结果实现头部的二次检测从而准确识别是否佩戴头盔，还可实现骑乘人员与周边行人的区分，并检测“一车多人”现象中是否佩戴头盔。实验表明，CSSD整体mAP为86.4%，FPS为47帧/秒。

针对CSSD进行改进，通过增加通道注意力机制以及SPP(Spatial Pyramid Pooling)网络，设计了一种引入视觉机制的VSSD(Visual in SSD)网络。可以提升小目标的检测性能，在深层网络中增加SPP可在扩大感受野的同时进行特征融合，改善大目标物体的识别。实验表明VSSD mAP达到94.97%，相比改进前的CSSD网络，mAP 提升8.57%，但FPS 稍有下降，为40帧/秒。

在网络压缩方面，首先将VSSD和CSSD作为教师与学生网络，通过知识蒸馏的方式让教师网络指导学生网络进行训练，从而在不改变网络结构的情况下提升学生网络的准确率。然后对蒸馏后的CSSD进行自蒸馏的“复习”阶段训练，自蒸馏后mAP达到95.8%。最后对自蒸馏后的CSSD主干部分进行通道剪枝，过滤掉贡献度小的通道，得到CSSD-P(CSSD Pruned)网络。实验表明，CSSD-P的mAP达到94.5%，同时与VSSD相比mAP几乎持平情况下FPS提升14帧/秒，使网络更加高效和精确，更好的满足检测任务。

At present, the supervision of helmet wearing in E-bike mainly relies on manual labor, which can cause waste of labor costs. About 80% of E-bike safety driving accidents in China result in brain injury and death. With the rapid development of intelligent security and the implementation of the "One Helmet and One Belt" policy proposed by the country, object detection technology has begun to be applied in the field of E-bike helmets. This article analyzes some classic object detection algorithms and selects SSD (Single Shot MultiBox Detector) as the basic algorithm for this study.

To solve the problem of helmet detection for E-bike, a cascade network CSSD that can realize secondary detection is designed on the basis of SSD, which takes the human and E-bike as a whole as a first detection, and then realizes the second detection of the head after the detection results, so as to accurately identify whether to wear the helmet, and also realize the distinction between the riders and the surrounding pedestrians, and detect whether to wear the helmet in the phenomenon of "one car with multiple people". The experiment shows that the CSSD overall mAP is 86.4%, and the FPS is 47 frames/second.

To improve CSSD, a VSSD (Visual in SSD) network with visual mechanism is designed by adding channel attention mechanism and SPP (Spatial Pyramid Pooling) network. It can improve the detection performance of small targets, add SPP in the deep network to expand the receptive field and perform feature fusion, and improve the recognition of large target objects. The experiment shows that the mAP reaches 94.97%. Compared with the improved CSSD network, the mAP increases by 8.57%, but the FPS slightly decreases to 40 frames/second.

In terms of network compression, VSSD and CSSD are used as teacher and student networks. Through knowledge distillation, the teacher network guides the student network to train, so as to improve the accuracy of the student without changing the network structure. Then the distilled CSSD was trained in the "review" stage of self-distillation, and the mAP after self-distillation reached 95.8%. Finally, channel pruning is performed on the backbone part of CSSD after self-distillation, and the channel with small contribution is filtered out to obtain CSSD-P (CSSD Pruned) network. The experiment shows that the mAP of CSSD-P reaches 94.5%. At the same time, compared with VSSD, the FPS is increased by 14 frames/second when the mAP is almost flat, which makes the network more efficient and accurate, and better meets the detection task.