The　escalating　complexity　of　urban　traffic　environments　demands　advanced　solutions　for　real-time　vehicle　detection　and　tracking.　This　paper　proposes　MAVIST　(Multi-stage　Adaptive　Vehicle　Identification,　Sensing,　and　Tracking　system),　an　integrated　framework　that　synergizes　enhanced　YOLOv5　object　detection　with　DeepSORT-based　tracking,　augmented　by　a　novel　Mosaic-8　data　augmentation　strategy　and　optimized　loss　functions.　By　partitioning　input　images　into　eight　subregions,　Mosaic-8　enriches　multi-scale　contextual　learning　and　mitigates　detection　ambiguities　caused　by　occlusions　and　scale　variations.　The　framework　incorporates　a　Complete　IoU　(CIoU)　loss　function　to　refine　bounding　box　regression　and　a　weighted　Focal　Loss　to　address　class　imbalance,　achieving　a　mean　average　precision　(mAP)　of　98.2%　on　urban　traffic　datasets.　Experimental　validation　demonstrates　MAVIST’s　capability　to　maintain　32　FPS　processing　speeds　while　accurately　tracking　vehicles　in　bidirectional　flows,　multi-directional　intersections,　and　regional　congestion　scenarios.　The　system’s　adaptive　anchor　optimization　and　attention-based　feature　fusion　enable　robust　performance　under　dynamic　lighting　conditions　and　dense　traffic　occlusions.　Case　studies　highlight　its　effectiveness　in　reducing　identity　switches　by　42%　compared　to　conventional　methods　and　providing　millimeter-level　trajectory　precision　for　congestion　analysis.　This　work　advances　intelligent　transportation　systems　by　offering　a　scalable,　data-driven　solution　for　real-time　traffic　monitoring　and　urban　mobility　optimization.　©　2025　SPIE.