Classification of criminal activities in surveillance footage using human detection and temporal sequence modeling.
| dc.contributor.author | Namuyomba, Angella Zerida | |
| dc.date.accessioned | 2026-09-12T11:40:45Z | |
| dc.date.available | 2026-09-12T11:40:45Z | |
| dc.date.issued | 2026 | |
| dc.description | Postgraduate Dissertation | |
| dc.description.abstract | Manual reviewing of surveillance footage to retrieve the evidence is a tiresome and time consuming activity due to the scarcity of frames containing these evidences. This study explores a framework for modeling spatial-temporal features to classify criminal activity in surveillance footage. This study proposes an approach that investigates the use of spatial features derived from YOLOv8 person detections as sequential representations of human behavior for LSTM-based crime classification in surveillance footage. We utilize deep learning object detection algorithms and Recurrent Neural Networks (RNNs) to learn the spatial-temporal features extracted from the UCF crime dataset in order to classify criminal activity in surveillance footage. Object detection allows us to focus our attention on frames that contain humans and human activity to avoid searching through empty frames. We believe that since most, if not all crimes involve human activity, it’s imperative to accurately detect people in the challenging surveillance footage. A comparative study was carried out to assess the performance of the common deep learning object detection tools i.e single shot mult-ibox detector (SSD) and you only look once version 8 (YOLOv8) in detection of humans in challenging surveillance footage. The dataset used for this study was compiled from the HR crime dataset which consists of 13 categories of crimes that are Abuse, Arrest, Arson, Assault, Road Accident, Burglary, Explosion, Fighting, Robbery,Shooting, Stealing, Shoplifting and Vandalism. The study indicated that the YOLOv8 object detection tool outperforms the SSD detector with a mean average precision (mAP) of 0.97 and intersection over union(IOU) of 0.89. The obtained bounding box information together with the confidence scores is converted into sequences which are then used to train and evaluate an LSTM model. The model attained an accuracy of 0.8091, F1 score of 0.8099 and MCC of 0.7919.. This implies that the model ably classfies the criminal activity in the surveillance footage and can facilitate users in faster acquisition of criminal evidence with minimal human supervision | |
| dc.description.sponsorship | Dr. Owomugisha Godliver ; Dr. Godfrey Mirondo Kibalya ; Busitema University | |
| dc.identifier.citation | Namuyomba, A. Z. (2026). Classification of criminal activities in surveillance footage using human detection and temporal sequence modeling. [Unpublished undergraduate research report]. Busitema University. | |
| dc.identifier.uri | https://bdears.busitema.ac.ug/handle/123456789/9454 | |
| dc.language.iso | en | |
| dc.publisher | Busitema University | |
| dc.title | Classification of criminal activities in surveillance footage using human detection and temporal sequence modeling. | |
| dc.type | Other |