Improvement of Object Detection Networks by Fusion of Attention Mechanism and Efficient Convolution*

Due to the special position of the camera in the elevator, the collected pictures present a top-down picture, and there is a certain degree of occlusion, which increases the difficulty of target detection. In this scenario, in order to avoid vehicles entering the elevator from the source, it is necessary to improve the accuracy and real-time of object detection. Aiming at the problems existing in this scenario, an object detection algorithm based on the improved YOLOv7 model is proposed. Based on the YOLOv7 detection network, part of the traditional convolution is replaced, the classification loss is optimized, image enhancement is turned on, and some Conv, Concat, and SPPCSPC in the network structure are combined with SE attention mechanism modules to improve the accuracy and speed of object detection. Experimental results show that compared with the original YOLOv7 algorithm, the improved algorithm increases map@0.5 from 80.4% to 89.8%, map@0.5:0.95 from 63.9% to 67.8%, and fps from 0.737 to 0.736 compared with the original YOLOv7 algorithm.