Integrated Recognition Approach and Fining System Towards Detection and Tracking of Traffic Violating Multiple Vehicles using Pairing Net and Light Weight Deep Sort Fast YOLO Rec Architectures

Innovations in intelligent transport management systems have enhanced vehicle anomaly detection through traffic monitoring sensors and computer vision. Current approaches, however, lack effectiveness in multi-vehicle detection and tracking amidst changing traffic, especially for overspeeding, overloading, helmet violations, frequent lane changes, and imitation or unrecognized license plates. This research study proposes an end-to-end recognition method and self-driving fining system based on PairingNet + Lightweight Deep SORT Fast YOLO-Rec Architecture in order to provide improved detection efficiency and speed. Real-time sensor data from benchmark datasets or live video streams is processed into the form of frames. The multi-scale features are extracted by Fast YOLO-Rec utilizing a CNN backbone with Cross-Stage Partial (CSP) Connection Blocks, Residual Blocks, and Dark Blocks. Path Aggregation Network (PANet) maintains precise detection for objects of multiple sizes. For tracking, an anchor box prediction technique produces bounding boxes, and PairingNet, which is a graph convolution network-based method, pairs vehicles between frames based on contour and texture similarity. Deep SORT algorithm classifies offenses like counterfeit license plates, helmet misuse, and overloaded vehicles. Alerts are triggered to authorities and vehicle owners when infractions are detected. Experimental results show that the suggested model performs better than traditional methods in detection accuracy and processing efficiency and is a stable solution for traffic law enforcement and road safety.