Video detection method of pointer instrument based on improved FCOS

Abstract

Aiming at the problem that pointer instrument detection algorithm has slow locating speed and low real time performance in edge equipment, this paper proposes a pointer instrument video detection method based on improved FCOS. The algorithm is based on FCOS model and uses lightweight network ShuffleNetV2 to extract image features. Using PAN structure to strengthen the original feature fusion network, a bidirectional feature fusion network is formed. The attention module with global context information is introduced to reduce the information attenuation in the process of feature fusion. The two parameters of pixel utilization PUR and relative time increase RIT are introduced to test the influence of images with different image pixels on the detection effect in a more intuitive form. Through experiments, when the resolution of the input image pixel is 1280×1280, compared with the baseline model, the detection time of the pointer instrument video detection method based on improved FCOS is reduced by 91.60% when the detection accuracy is similar.