Improving multi-object re-identification at night with GAN data augmentation

This study concentrates on a camera-based traffic sensor that measures bicycle, vehicle and pedestrian trips called FlowCube™. To achieve multi-object tracking, FlowCube uses a model chain consisting of object detection, local tracking, trip filtering and re-identification (re-id). Whereas FlowCube’s performance is fit-for-purpose during the daytime, it degrades in more challenging nighttime conditions. With that, this study is aimed at improving FlowCube’s nighttime re-id performance. The hypothesis is that the poor nighttime re-id performance is due to a lack of nighttime re-id training data. So, in this paper a Generative Adverserial Network based data augmentation with alpha blending is proposed to enrich FlowCube’s re-id training data with synthetic nighttime imagery. The findings show that this method improves FlowCube’s mean re-id F1 scores and reduces the variance between results across multiple training runs, both for nighttime and general re-id. The same improvement can be expected for other camera-based traffic sensors which use multi-object tracking with re-identification.