NUST Institutions Library Catalogue

Smart city development places high emphasis on high-performance safety solutions to
counter city security threats such as street crime and increased numbers of unconcealed
and concealed objects. Although unconcealed objects are visible threats that are relatively
easy to identify, their multi-class counterparts that are concealed pose high challenges
given that they are not visible, especially within dense thermal grayscale scenes. Prevalent
detection solutions like metal detectors and infrared cameras usually fail to correctly
recognize multi-class objects, either unconcealed or concealed objects, which often result
in false negatives and require human verification. This research presents an enhanced
multiclass object detection model using YOLOv9 combined with Convolutional Block
Attention Modules (CBAM) to improve improvement of inference time with a special
emphasis on detecting concealed objects. The model embeds channel and spatial attention
from CBAM within YOLOv9’s backbone and head to enhance feature representation for
multi-scale object detection and efficiently handle unconcealed as well as concealed
objects from thermal images. The model applies YOLOv9’s Programmable Gradient
Information (PGI) and Generalized Efficient Layer Aggregation Network (GELAN) to
streamline feature extraction and reduce loss of information with enhanced multi-target
detection accuracy. The model was trained and tested on a set of 8,600 grayscale image
augmentations, spanning four object classes (revolvers, tasers, knives, and guns), with
strict preprocessing to keep classes balanced and semantically sound. The YOLOv9-
CBAM structure obtains a mean Average Precision (mAP@0.5) of 0.954 and mAP@0.5-
0.95 of 0.948 compared to the baseline YOLOv9's mAP@0.5 of 0.944 and mAP@0.5-0.95
of 0.940, representing accuracy improvements. The CBAM-augmented model achieves
1.9% faster inference speed than the baseline, achieving optimal performance with both
unconcealed and hidden object detection at optimized computation efficiency. The efficient
and highly effective system best suits real-time smart city security applications like airport
screening, patrol by police, and critical infrastructure security, reducing queuing time
considerably while significantly enhancing automated threat detection efficiency.