Use this URL to cite or link to this record in EThOS: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.768075
Title: Applying computer analysis to detect and predict violent crime during night time economy hours
Author: Lloyd, Kaelon
ISNI:       0000 0004 7652 3545
Awarding Body: Cardiff University
Current Institution: Cardiff University
Date of Award: 2018
Availability of Full Text:
Access from EThOS:
Access from Institution:
Abstract:
The Night-Time Economy is characterised by increased levels of drunkenness, disorderly behaviour and assault-related injury. The annual cost associated with violent incidents is approximately £14 billion, with the cost of violence with injury costing approximately 6.6 times more than violence without injury. The severity of an injury can be reduced by intervening in the incident as soon as possible. Both understanding where violence occurs and detecting incidents can result in quicker intervention through effective police resource deployment. Current systems of detection use human operators whose detection ability is poor in typical surveillance environments. This is used as motivation for the development of computer vision-based detection systems. Alternatively, a predictive model can estimate where violence is likely to occur to help law enforcement with the tactical deployment of resources. Many studies have simulated pedestrian movement through an environment to inform environmental design to minimise negative outcomes. For the main contributions of this thesis, computer vision analysis and agent-based modelling are utilised to develop methods for the detection and prediction of violent behaviour respectively. Two methods of violent behaviour detection from video data are presented. Treating violence detection as a classification task, each method reports state-of-the-art classification performance and real-time performance. The first method targets crowd violence by encoding crowd motion using temporal summaries of Grey Level Co-occurrence Matrix (GLCM) derived features. The second method, aimed at detecting one-on-one violence, operates by locating and subsequently describing regions of interest based on motion characteristics associated with violent behaviour. Justified using existing literature, the characteristics are high acceleration, non-linear movement and convergent motion. Each violence detection method is used to evaluate the intrinsic properties of violent behaviour. We demonstrate issues associated with violent behaviour datasets by showing that state-of-the-art classification is achievable by exploiting data bias, highlighting potential failure points for feature representation learning schemes. Using agent-based modelling techniques and regression analysis, we discovered that including the effects of alcohol when simulating behaviour within city centre environments produces a more accurate model for predicting violent behaviour.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.768075  DOI: Not available
Share: