As　autonomous　driving　technology　advances,　real-time　hazard　detection　is　becoming　critical　for　safety.　Convolutional　Neural　Networks　(CNNs),　known　for　their　powerful　image　processing　capabilities,　play　a　key　role　in　autonomous　vehicle　perception　systems,　excelling　in　tasks　such　as　object　detection,　recognition,　and　classification.　However,　existing　CNN　models　face　a　trade-off　between　accuracy　and　processing　speed,　particularly　in　complex　traffic　scenarios.　This　research　proposes　a　novel　CNN　architecture　paired　with　a　dedicated　hardware　accelerator　to　achieve　high-precision　hazard　detection　without　compromising　real-time　performance.　The　model　features　an　attention-based　dynamic　CNN　structure　that　adjusts　convolutional　kernel　sizes　and　quantities　based　on　image　complexity,　optimizing　resource　allocation.　A　cross-layer　feature　fusion　strategy　improves　detection　of　small　and　distant　hazards　like　pedestrians　and　obstacles.　Additionally,　a　custom　AI　accelerator　using　FPGA　technology　incorporates　sparse　matrix　operations　and　dynamic　data　flow　management　to　enhance　inference　efficiency　and　reduce　power　consumption.　Experimental　results　in　real-world　traffic　environments　demonstrate　high　accuracy,　low　latency,　and　energy　efficiency,　confirming　the　system＇s　robustness　and　potential　for　deployment　in　autonomous　vehicles.　©　2025　IEEE.