基于集成学习的红外双波段被动测距模型

Passive IR Ranging Based on Ensemble Learning

摘要: 为了提高红外被动测距模型的精度和实用性，基于集成学习的方法，提出了一种由数据驱动的红外被动测距模型。首先利用MODTRAN生成3.5 um~4.0 um和4.5 um~4.7 um波段下不同距离和高度角的大气透过率数据，然后结合红外辐射传输理论将这些数据转换为红外探测器对应的电压值，接着在集成学习的boosting框架下，利用这些电压数据分别建立gbdt、xgboost、adaboost测距模型，并对模型参数进行了优化，然后采用集成学习的stacking方法对上面建立的3种模型进行融合作为最终模型，最后用测试数据检验了该模型的预测效果，为了作对比，论文还使用了average融合方法。结果表明，采用stacking融合的模型能够实现很高精度的预测。

Abstract: To improve the practicability and accuracy of passive IR ranging, a new passive IR ranging method based on ensemble learning is proposed. The atmospheric transmittance data of 3.5 um - 4.0 um and 4.5 um - 4.7 um in different range are generated by MODTRAN, then these data are converted to voltage based on infrared transmission theory, using these voltages to establish single gbdt, xgboost and adaboost models under the framework of boosting, and then merge the three models with stacking method as the final model. At last, the accuracy of the model is tested by testing data. For comparison, average method is also used to merge the three models above. The result shows that stacking method can make further improvement of accuracy.