Motion Embedding for On-road Motion Object Detection for Intelligent Vehicle Systems

Accurate motion object detection (MOD) using in-vehicle cameras in driving vehicles is a challenging task. Several deep learning based motion segmentation approaches have been reported based on the interpretable optical flow feature. However, the interpretable optical flow feature has not been explored by object-level MOD approaches. In this paper, we propose a motion embedding pipeline (MEP) architecture that utilizes interpretable optical flow and deep learning to solve object-level MOD problems. The MEP is a three-stage pipeline that consists of an object detector, a novel feature extraction algorithm to capture relative motion between objects and the background, as well as a motion predictor for representation learning with stacked autoencoder to determine motions. A new dataset, Singapore motion object detection (SG-MOD) dataset is constructed in this work with much larger variations in urban environments. Experimental results show that the proposed MEP outperforms other pipeline-based architecture and deep learning based approaches on the SG-MOD and KITTI-MOD datasets in most metrics.