AggNet: Simple Aggregated Network for Real-Time Multiple Object Detection in Road Driving Scene

A real-time object detection model in road driving scene is being established as an important component technology that facilitates autonomous driving; its performance has been improving drastically due to progress in deep learning technology. However, the performances for occlusion between objects and small object detection are not yet perfect. In this paper, a simple aggregated convolution neural network (AggNet) based real-time multiple object detection model is proposed to improve the detection performance of occlusion or small objects in road driving scene. The proposed model is designed to deliver feature information of a small receptive field by improving the residual block of the traditional Residual Network (ResNet) with the simple aggregation block. Furthermore, to detect objects of various sizes effectively, an aggregated rezoom layer-based object detection method was applied instead of the conventional multi-scale feature and anchor-based object detection method. As a result of tests using KITTI data, it was confirmed that the proposed model can successfully detect small objects and object occlusions of various sizes. 1024x320 input images can be processed at 20-22 fps with an NVIDIA Titan XP GPU.