Real-time Jet Segmentation Based on Mixed Dilated Convolution MobileNetV2
In order to quickly and accurately extract the jet trajectory so that the smart water cannon can automatically adjust the strike direction according to the target movement during the strike,a real-time jet segmentation model based on the mixed dilat-ed convolution MobileNetV2 is proposed.The model uses the lightweight network MobileNetV2 as the basic network,and only uses the first 7 layers of the network to achieve fast real-time segmentation of jet trajectories,a mixed cavity convolution layer is added to collect multi-scale semantic information of different receptive fields to improve jet trajectory segmentation accuracy.Finally,com-bine coding the decoder-decoder structure links the output of the MobileNetV2 network and the output of the mixed-hole convolu-tional layer to further refine the segmentation results.By building a small intelligent water monitor platform and collecting data sets for experiments,the experimental results show that the average intersection ratio of the model reaches 0.781 6,and the average speed of dividing a frame of picture is 19 ms,and it has achieved good performance in segmentation accuracy and segmentation speed.
MobileNetV2dilated convolutionencoder decoderreal time segmentationintelligent water cannon
万方数据
维普
