FED: A Method for Detecting Indoor People Looking down under a Fisheye Lens
Abstract
The Faster Region-based Convolutional Neural Network (Faster R-CNN) object detection model has become a milestone algorithm in deep learning due to its high precision and low computational complexity, but it still has certain limitations. In order to be suitable for indoor people detection under the fisheye lens, this paper proposes a fast object detection method, Fisheyes Effective Detection (FED) based on Faster R-CNN. First, Ameliorate Intersection over Union (AIoU) is used to improve the accuracy of the positioning phase. Secondly, using the ResNet with Residual-I module and adding the channel attention mechanism to extract more useful object feature information and reduce detection errors. Finally, when constructing the network, the FED detection network is constructed by changing the stacking order and replacing the activation function. Experimental results show that FED has superior detection performance, faster detection speed, and object recognition performance. It is more suitable for indoor personnel detection under a fisheye lens.