Multi-scale Object Detection Algorithm in Smart City Based on Mixed Dilated Convolution Pyramid

Abstract

In the current smart city scene, the positioning, detection and tracking of the objects are essential technologies in urban security. As the most prominent problem in the field of object detection, occlusion and multi-scale seriously affect the recall and accuracy of the algorithm. In response to the above problems, this paper starts from the receptive field and proposes an object detector based on the Mixed Dilated Convolution Pyramid Network (MDCPN). First of all, the dilated convolutional layers of different sizes are introduced into the feature pyramid to construct a Mixed Receptive Field Module (MRFM), which aims to obtain more global feature information by increasing the receptive field under the condition of controlling the amount of parameters, then solve the problem of occlusion of the object; secondly, improve the structure of the feature pyramid, and design a Lower Embedding Feature Pyramid Module (LEFPM), which combines low-level feature detail information and high-level feature semantic information to improve the representation ability of feature maps, and enhance the model scale adaptability; in particular, for the problem of missed detection, the anchor free mechanism of the FCOS algorithm is introduced, which effectively reduces the redundancy of candidate frames and further improves the positioning accuracy. Tested on the public dataset VOC, the algorithm has greatly improved the detection accuracy.