Helmet Wearing Detection Based on YOLOv4-MT
Jie Liu, Lizhi Liu · 2021
Current networks based on methods such as YOLO, SSD and R-CNN are more complex and large, making it difficult to use on embedded devices without high-performance Gpu support. Aiming at the application of the model in embedded devices, an improved YOLOv4-MT helmet wearing detection model is proposed to keep the detection speed and accuracy basically unchanged while reducing the model volume. YOLOv4-MT combines the advantages of MobileNet and YOLOv4, and uses K-means algorithm to cluster the prior box to improve the network output scale. The backbone module of YOLOv4 is replaced by the deep separable convolution module of MobileNet. The head part of YOLOv4 is retained and improved, and a new residual module is built in the output part. The CIOU loss function is used to evaluate the loss function. The pre-training model is loaded, and the freezing training is used to freeze part of the weights and train the network parameters of the latter part. Self-made data set based on construction site background, including a variety of complex construction environment. In order to verify the detection effect and performance of the model, several models are compared in NVIDIA GeForce GTX 1650 experimental environment. The new model YOLOv4-MT improved by 7% compared with YOLOv4-Tiny mAP, and reduced by 7 times compared with YOLOv4 model size. Compared with MobileNet-SSD, the detection speed is 4.6 times higher.