Object Detection Using a Single Extended Feature Map

Abstract

Fully convolutional neural network-based object detectors have achieved considerable detection accuracy in recent years. It is a recent trend to establish complex and deep network architectures for improvement of the detection accuracy. However, object detectors for intelligent vehicle applications require fast inference speed, Iightweight network architecture, and less memory usage as well as high detection accuracy to implement the algorithm in an embedded hardware. In this paper, we propose a fast object detection method based on a single stage and a single extended feature map. A Iightweight network based on an extended paththrough layer is proposed to improve both the accuracy and speed. The extended paththrough layer enlarges the resolution of the last feature map by concatenating later feature maps with lower resolution to earlier feature map maps with higher resolution. The layer helps to search and detect smaller objects more densely on the extended last feature map. Our experimental results show that the proposed detection model outperforms the previous state-of-the-art methods in both detection accuracy and inference speed. © 2018 IEEE.