Multi Task Learning Architecture for Vehicle Detection and Vehicle Tracking Towards Passenger Safety and Traffic Violations Detection Using Pairing Net and Fast Yolo Rec Approach

• Published in: 2025 International Conference on Machine Learning and Autonomous Systems (ICMLAS) pp. 1345 - 1349
• Main Authors: R, Mugesh, R, Manoj, R, Kaviprasth, S, Gokilavani
• Format: Conference Proceeding
• Language: English
• Published: IEEE 10.03.2025
• Subjects: Accuracy; Automated Fining; Convolution; Convolution Neural Network; Fast Yolo Rec Algorithm; Feature extraction; Helmet use Classification; Helmets; License plate recognition; Monitoring; Neural networks; Number of rider detection; Predictive models; Road safety; Safety; Vehicle Classification; Vehicle Detection; Vehicle Monitoring
• DOI: 10.1109/ICMLAS64557.2025.10967612

Nowadays, Traffic violation is becoming serious challenges to the government and insurance agencies as it results danger to their life or other life's. Especially two-wheeler vehicle leads to primary and big concern against traffic violations involving bike racing, carrying more than two passengers, driving without helmets, lane changes. Many manual fining system has been imposed against violators against the road safety still many incident occurring during absence of the traffic monitoring polices. In order to prevent the traffic violations automatically, high strength traffic violation monitoring and detection system has to be integrated with automatic fining system. In this paper, multitask learning architecture for monitoring, detection and preventing the two-wheeler vehicle passenger safety and traffic violations is developed along automated violation fine ticketing system using Graph Convolution l Network and Fast Yolo Rec techniques. Proposed architecture is capable of detecting the persons not wearing the helmet, wearing helmet improperly, bike racing, lane changing and triple ridding to impose the automatic traffic and safety violation fine ticketing on identifying vehicle number plate details using number plate recognition system. Initially proposed architecture handles the acquired video from the real monitoring system or extracting video from coco dataset which is considered as benchmark dataset for intelligent transportation systems developments. Extracted video is transformed into image frames. Those Image frames were processed using convolution Neural Network for segmentation and feature extraction on the convolution and fully connected layers of the model and those extracted feature were employed to detection using Fast Yolo Rec approach. Fast Yolo Rec approach is object recognition mechanism which is capable of detecting multiple objects in a single frame within short period. Fast Yolo Rec is capable of producing bounding box around the object of interest. Fast Yolo Rec architecture composed of ResNet as backbone which contains CNN layers with receptive fields, PANet as path aggregation neck and head. ResNet is employed to process the extracted the feature and map those feature related to number plate and no of riders to the specified vehicles in the particular frame. However, neck is utilized for feature mapping of various positions related to helmet usage and no of riders. Finally, head is utilized for detection of helmet use and no of passenger in the particular vehicle. Automated ticket fining is carried out on detection of abnormality on the specified vehicle. Proposed architecture is experimented and evaluated with respect to detection accuracy, average precision and frames per second on detecting the no of riders, helmet use of monitored vehicle. Experimental results prove that proposed architecture provides better results compared to conventional approaches on accuracy and efficiency on processing road safety related information.