Robust Pedestrian Detection via Curated Training on Created Dataset
Sukesh Babu V S, Rahul Raman · 2024
Pedestrian detection is a challenging research area in computer vision with various applications, viz. autonomous driving, visual surveillance, walk assistance, and traffic monitoring. Identifying pedestrians accurately is a difficult task involving detecting pedestrians in different lighting conditions, occluded pedestrians, multi-scale detection, crowded environments with varying poses, and human-like objects. Several pedestrian datasets are available, but none contain all the variations that will happen in the real-time environment. Using a pretrained model for pedestrian detection often results in poor results when tested on real-time data. To address this issue, we created a pedestrian dataset (CamPed: Campus Pedestrians: https://github.com/RahulRaman2/CamPed-Dataset) with 100K images, with a total of 400K pedestrian instances, including most of the diversities in real-world environments. The dataset is trained and tested on YOLO, DETR, and RT-DETR detection models. These models were also trained and tested on standard pedestrian datasets WiderPerson, PennFudan, INRIA, and PnPLO, and the results were compared with those obtained when training and testing with our dataset. The results show that the models trained on the CamPed dataset give better mAP (mean Average Precision) than those trained on another pedestrian dataset. We have tested the models trained on our CamPed dataset on several standard pedestrian datasets, and the experimental results show that the models trained on our dataset improve the detection accuracy of small pedestrian blobs, detection in crowded environments, and pedestrians with different poses. At the same time, the custom-trained models give better rejection to pedestrian-like objects such as human cutouts, effigies, figurines, or sculptures.