ENHANCED YOLOV5-BASED MODEL FOR REAL-TIME DETECTION OF SMOKING AND PHONE USE IN DANGEROUS DRIVING
Abstract
This research focuses on the automatic detection of two perilous driving behaviors: smokingandengaginginphoneconversations.Atfirst,wecreatespecializeddatasetsfortheseabnormalacts.Inaddition,wepresentatwo-stageapproachforidentifyingbehavioralirregularities. The efficacy of the YoloV5 object detection network model in detecting smalltarget objects is improved through the strengthening of its prediction head. We optimize theloss function for datasets that consist of categories that are mutually exclusive. The structureof the posture estimation network is enhanced by integrating the attention mechanism of theCoordinate Attention (CA) structure. This integration aims to improve the efficiency andaccuracy of information processing. The assessment of the final outcome is conducted byemploying Euclidean distance calculation, with the elbow joint angle acting as an additionalcriterion for judgment. The suggested model for identifying hazardous driving behaviorsachieves a mean average precision of 93.4% at a speed of about 61 frames per second (FPS),resulting in an improvement of 8.2% in detection accuracy. This fulfills real-time demandsandenhances precision while maintainingvelocity.
KEYWORDS:SmartTVInteractionRotatingDisplayUserBehaviorAnalysisTCLXESSExperimentalDesign