Integration of Object Detection in Crime Scene Investigation
Issued Date
2025-01-01
Resource Type
eISSN
27304175
Scopus ID
2-s2.0-85213388789
Journal Title
Engineering Access
Volume
11
Issue
1
Start Page
13
End Page
23
Rights Holder(s)
SCOPUS
Bibliographic Citation
Engineering Access Vol.11 No.1 (2025) , 13-23
Suggested Citation
Werukanjana P., Permpool N., Sa-Nga-Ngam P. Integration of Object Detection in Crime Scene Investigation. Engineering Access Vol.11 No.1 (2025) , 13-23. 23. doi:10.14456/mijet.2025.2 Retrieved from: https://repository.li.mahidol.ac.th/handle/20.500.14594/102822
Title
Integration of Object Detection in Crime Scene Investigation
Author(s)
Author's Affiliation
Corresponding Author(s)
Other Contributor(s)
Abstract
Object detection is applicable in various fields, encompassing Crime Scene Investigation. The act of capturing evidence through photography in crime scenes is of paramount importance. Concurrently, object detection and scene analysis can be integrated by the investigator during this process. Nonetheless, the investigative procedure involves multiple phases. In this research, object detection is employed to identify crucial evidence discovered at crime scenes, along with objects commonly linked to assault cases as statistics reported by the National Institute of Justice (NIJ) in the United States of America. Challenges such as pen guns and illegal firearms persist in criminal activities, although they are not no category of legal firearms. Transfer learning techniques are adopted in this study from an existing model, utilizing pre-trained models to identify crucial evidence in crime scenes. The experiment gathered 494 images of pen guns, illegal firearms, and associated objects from online police news reports to form a custom dataset. The process of annotating, training, evaluating, and fine-tuning the custom dataset led to experimental outcomes demonstrating a high Mean Average Precision (mAP) across all target’s custom datasets. The model reached convergence at approximately epoch 100, achieving high precision as 0.97, recall as 0.926 and box mAP50 of as 0.974 Additionally, box mAP50-95 as 0.817 However, this paper presents a confusion matrix of customized specific classes with a low volume dataset. These findings highlight the potential application of object detection in crime scene investigations with the aid of object detection. Consequently, this approach could aid in formulating a project blueprint for a Crime Scene Investigation model and furthering the detection of evidence objects in the future with very large volume of dataset. In conclusion, this study illustrates the capability of detecting pen guns, illegal firearms, and related objects through object detection techniques.