Real-time spectral imaging of dynamic scenes is critical for industrial safety monitoring， atmospheric remote sensing， and aerial target recognition. Traditional spectral imaging methods are constrained by sequential sampling and low optical throughput， limiting simultaneous achievement of high temporal resolution and high detection sensitivity. Computational spectral imaging offers a promising solution， yet LWIR implementations remain challenging due to difficulties in coding element fabrication， response matrix calibration， and training data scarcity. A LWIR computational video spectral imaging method based on a broadband random coding array was developed， addressing coding element fabrication， system calibration， and dataset construction. Using 9 broadband randomly coded channels， 32 band spectral data cubes were reconstructed at 200 nm spectral resolution and 30 Hz frame rate， with detection sensitivity 5–10 times higher than equivalent narrowband filter schemes. In a gas leakage scenario， 16 gas species were identified with 98.97% average accuracy using a characteristic absorption spectral library. These results confirm the potential of this approach for gas spectral measurement and recognition in real dynamic scenes.