SSD-ML: hierarchical object classification for traffic surveillance

We propose a novel CNN detection system with hierarchical classification for traffic object surveillance. The detector is based on the Single-Shot multibox Detector (SSD) and inspired by the hierarchical classification used in the YOLO9000 detector. We separate localization and classification during training, by introducing a novel loss term that handles hierarchical classification. This allows combining multiple datasets at different levels of detail with respect to the label definitions and improves localization performance with non-overlapping labels. We experiment with this novel traffic object detector and combine the public UADETRAC, MIO-TCD datasets and our newly introduced surveillance dataset with non-overlapping class definitions. The proposed SSD-ML detector obtains 96:4% mAP in localization performance, outperforming default SSD with 5:9%. For this improvement, we additionally introduce a specific hard-negative mining method. The effect of incrementally adding more datasets reveals that the best performance is obtained when training with all datasets combined (we use a separate test set). By adding hierarchical classification, the average classification performance increases with 1:4% to 78:6% mAP. This positive result is based on combining all datasets, although label inconsistencies occur in the additional training data. In addition, the final system can recognize the novel ‘van’ class that is not present in the original training data.