CGMA: An improved multi-attribute CIoU-guided enabled pedestrian detection

The study of pedestrian detection is essential for improving safety in urban and automobile environments, enhancing security and surveillance, assisting with urban planning, and fostering technological advancements. The elimination of redundant boxes is a crucial step in pedestrian detection, commonly accomplished through the utilization of Non-Maximum Suppression (NMS). Nevertheless, in crowded environments, the identification of occluded pedestrian becomes challenging due to their similarity to duplicate proposals. Consequently, this issue results in inaccurate detection outcomes and missed detections. To tackle the aforementioned issue and, a novel improved multi-attribute CIoU-guided enabled pedestrian detector is proposed. This detector incorporates attribute maps into the CGMA-NMS algorithm, which operates on multiple thresholds (hard and soft) based on dense and diverse attributes. By doing so, it effectively distinguishes between pedestrians that are overlapped with each other and duplicate ones, leading to the refinement of bounding boxes. Consequently, this approach enhances the accuracy of pedestrian detection in crowded conditions. Additional losses, as well as the CIoU loss, are incorporated into the detection heads, and a comprehensive loss function is introduced. This function significantly reduces the occurrence of missed detection and enhances the overall accuracy level. This approach is evaluated on a custom dataset i.e. GEUCampusTour and two other standard datasets CityPersons and CrowdHuman, further comparative analysis with the state-of-the-art is done between them. GEUCampusTour achieved recall of 94.4% and average precision (AP) of 90.2%, CityPersons achieved 8.1%, 43.0%, 8.1%, 5.2% of log average miss rate on reasonable, heavy, partial and bare occlusion respectively and CrowdHuman achieved 96.8%, 90.5%, 36.7% of recall, AP and MR⁻² respectively.