Improve Object Detections in Traffic Scenes

Abstract

Transportation is one of the largest area that can benefit from actionable insights derived from traffic image data. Object detections in traffic scenes aim to extract accurate localization and classification information. Since recognition and tracking often rely on the results from detection, a fast and accurate object detection framework can make transportation systems safer and smarter. However, there are several major road blocks for object detection in traffic scenes when training model in real world traffic dataset: (1) state-of-the-art deep neural network relies on large and well labeled supervised data (2) far away and crowded objects are hard to precisely located (3) quality of crowd-sourcing annotation is no guaranteed (4) some categories only have few training instances for training. To address the issues, a multi-window method merging with pre-trained model is proposed. Experiments on NVIDIA AI City Dataset shows that the proposed method has 12 to 57 % of average precision improvement for categories with few training instances and 3% of mean average precision improvement. It is more clear to understand the improvement through the qualitative results in section 3.3.