Swin transformer based vehicle detection in undisciplined traffic environment

Intelligent vehicle detection (IVD) plays a prominent role in evolving an intelligent traffic management system (ITMS). It can help to decrease the average waiting time at the traffic post, save fuel consumption, control traffic congestion, decrease accident rates, and build up human safety. Recent developments in the artificial intelligence (AI) domain have increased the demand for IVD in the undisciplined traffic environment, which is a usual condition in developing countries. IVD is a difficult task in an undisciplined traffic environment because different vehicle categories travel very close to each other on the roads and do not follow traffic rules. Previously, several convolutional neural network (CNN) based deep learning (DL), and visual transformer-based techniques for vehicle and object detection have been presented. They are complex and do not accurately extract multi-scale features due to the involvement of existing CNN feature extraction backbones. Also, most techniques failed to account for an undisciplined traffic environment due to the unavailability of labeled vehicle datasets. Therefore, this paper proposes a swin transformer-based vehicle detection (STVD) framework in an undisciplined traffic environment. Swin transformer (ST) wholly exchanges information within and between image patches and provides hierarchical feature maps, effectively alleviating the multi-scale feature extraction problem. A bi-directional feature pyramid network (BIFPN) is presented, which combines low-resolution features with high-resolution features in a bidirectional way and provides robust multi-scale features with different scales and resolutions. A fully connected vehicle detection head (FCVDH) is applied to improve the matching relationship between vehicle sizes and the BIFPN hierarchy. FCVDH predicts the locations and categories of vehicles in the input image. STVD is analyzed, experimented, and measured over realistic traffic data. Also, it is compared with the existing state-of-the-art vehicle detection methods. It achieves 91.32% detection accuracy on diverse traffic labeled dataset (DTLD), 87.4% on IITM-hetra, and 88.45% on KITTI datasets.