TIRNet: Object detection in thermal infrared images for autonomous driving

In the present study, towards reliable and efficient object detection in thermal infrared (TIR) images, we put forward a novel object detection approach, termed TIRNet, which is built upon convolutional neural network (CNN). Instead of using the deep CNN backbone (ResNet, ResNeXt) which suffers low speed and high computational cost, the lightweight feature extractor (VGG) is adopted. To get the robust and discriminating features for accurate box regression and classification, the Residual Branch is introduced. More uniquely, it only exists in the training phase, so no any additional time is increased when inference. All the computation is encapsulated in a single network, so our TIRNet can be optimized and tested in the manner of end-to-end. Furthermore, the continuous information fusion strategy is proposed for improving detection performance, which can effectively solve the problems such as complex background, occlusion, and get more accurate and smoother detection results. To get the real-world dataset and effectively evaluate the effectiveness, a China Thermal Infrared (CTIR) dataset is collected. Besides, we also evaluate our proposed approach on the public KAIST Multispectral dataset. As demonstrated in the comparative experiments, our approach gets the state-of-the-art detection accuracy while maintains high detection efficiency.