Enhancing crime prediction with machine vision: Convolutional neural network approach

Abstract

Crime significantly challenges socio-economic development, hindering societal progress and stability. This PhD Thesis focuses on training and comparing YOLOv4 with SSD models for classifying high-risk perpetrators. The annotated training dataset comprises approximately 3,118 images, including common weapons such as shotguns, handguns, rifles, and knives. It was sourced from the UCF and the NIST criminal databases. The secondary sources are trusted, compliant with data privacy laws, and personal identifiable information was anonymized before used for training. The models’ validation shows effective object detection, with YOLOv4 achieving a mean Average Precision (mAP) of 90.58%. It achieves a 75.99 % precision for identifying persons of interest between 4,000 and 10,000 epochs. This model excelled most at detecting knives, achieving a precision of 96.11% with 80 instances predicted correctly (TP) with 2 incorrect prediction (FP). The recall rate of 74.5% for YOLOv4 indicates the model identified about 74.5% of actual positive instances. Moreover, with an F1 score of 77.2%, our findings highlight a balance between precision and recall. However, there is room for improvement due to 259 (FN) missed objects out of the 1205 total predictions. In contrast, the SSD model achieved 66.5% precision for persons, with its highest precision for handguns at 81.42 %. But adjusting configuration variables (hyperparameter tuning) improved the model’s mAP to 84.19%, indicating better generalization and prediction across classes. Both architectures performed well by reducing misclassification including maintaining relatively high true positive and low false positive rates across all classes. To enhance pre-crime analysis, the YOLOv4 model was integrated with the Deep-SORT tracker to maintain object identities over time and improved the model's ability to identify weapon sub-objects on individuals and categorize potential perpetrator as high-risk.