Lightweight Bird Eye View Detection Network with Bridge Block Based on YOLOv5

Abstract

In this paper, The network with a faster detection speed than the original YOLOv5 nano model is proposed. The network defined as a bridge module reduced the number of channels and changed the speed quickly by applying pixel-wise operation instead of using a convolution layer. Especially, element-wise addition operation of each output feature maps is the main method. As a result, the detection speed is faster than the original detection method about 30 35%. On the other hand, mAP (mean average precision) is recorded at 50.7%, which is 1.4% lower than the original detection method. However, the original detection method showed good results in 3 classes and the proposed method showed good results in 5 classes. And the proposed method detected more objects in a detection result image. Therefore, the proposed method is a more efficient object detection network.