Nighttime Vehicle Detection Algorithm Based on Improved Faster-RCNN

Abstract

Vehicle detection is important for the development of Intelligent Transportation Systems (ITS), which has made great strides in recent years. However, at night, vehicle detection faces many difficulties such as low illumination, street lights, and the appearances vehicle headlights, etc. In order to solve these problems, we propose an improved nighttime vehicle detection algorithm based on Faster R-CNN. Firstly, we combine the Deformable Convolutional Network with Faster R-CNN to improve the detection accuracy features of night vehicles of different sizes and shapes. Secondly, to improve the prediction accuracy of bounding box position information, we adopt Side-Aware Boundary Localization to replace the traditional bounding box prediction. It can further obtain more accurate position information. At the same time, aiming at the imbalance of samples in the training process, we use Oline Hard Example Mining(OHEM) to train samples with a high probability of error to improve the learning effect of a few classes; and to improve the accuracy of night vehicle detection. In addition, we use Soft Non-Maximum Suppression(Soft-NMS) to reduce the number of missed vehicles. The improved algorithm efficiently improves the night vehicle detection accuracy and reduces the model complexity. Furthermore, we verify the effectiveness of each innovation module through ablation experiments and comparison experiments. Finally, the advantages of the improved model in terms of nighttime vehicle detection accuracy are verified by experimenting on the open-source intelligent traffic dataset UA-DETRAC and the open-source diverse automated driving dataset BDD100K.