Object detection for panoramic images based on MS‐RPN structure in traffic road scenes

Abstract

The objection detection of panoramic image is the key part of street view, intelligent transportation, automatic driving and other technologies. Due to the shortcomings of existing algorithms in detecting panoramic images, firstly a high‐resolution panoramic image dataset is introduced, then the multi‐scale feature pyramid networks (MS‐RPN) structure is proposed and a new network with Sim‐Inception module is designed. The network can extract different scales of objects from different feature layers, so that the small object in the image can also be accurately detected. Finally, the entire detection network is trained by using the dataset constructed in this study. Meanwhile, the ROIPool is replaced by ROIAlign and the loss function is adjusted according to the network structure. The experimental results show that the detection performance on the panoramic dataset is significantly improved by authors’ proposed algorithm, which is better than other deep learning algorithms, especially for small object in the image.

Keywords

• MS-RPN structure
• traffic road scenes
• objection detection
• high-resolution panoramic image dataset
• multiscale feature pyramid networks structure
• different feature layers