Improved Faster RCNN-based Nighttime Pedestrian Detection Using RGB Images

The detection of pedestrians under less illumination or darker conditions is a key challenge in computer vision and is essential to many applications in the real world including robotics, surveillance, and self-driving cars. For the development of reliable autonomous systems, the performance of the current method for identifying pedestrians must be greatly improved. This work proposes a framework for nighttime pedestrian detection based on Faster RCNN to increase the dependability of features acquired from images taken in the nighttime (RGB images). To optimize the feature learning pipeline, the proposed architecture explicitly integrates a Transformer module and shuffle attention (SA) techniques. Shuffle Attention mechanism integrated with Bottleneck layer to make more attention to the pedestrian region rather than background region. Next Transformer block added in the Backbone to recognize the local and global features that the image possesses. The RGB night time images are curated from the benchmark datasets (ECP, BDD100100k, NightOwl) to represent as much variability with a wide range of light conditions. To analyse the proposed model performance tested with night time images. The experimental findings show that the improved Faster RCNN model raises the average detection accuracy (mAP@50) by 3.4% compared with baseline models. The detection result shows that model had good ability to detect pedestrian at poor light.