视觉车辆识别迁移学习算法

Vision based vehicle detection transfer learning algorithm

针对采用大样本离线训练的车辆识别分类器在新场景中性能显著下降的问题,提出了一种具有样本自标注能力的车辆识别迁移学习算法,并采用概率神经网络(probability neural netw ork,PNN)进行分类器训练.首先,提出一种基于多细节先验信息的样本标注策略,融合复杂度、垂直平面和相对速度等先验信息实现新样本的自动标注;然后,充分利用PNN训练速度快以及增加新样本时只需分类器进行局部更新的特点,将其引入到分类器训练模型中,取代传统机器学习算法中的Adaboost分类器.实验结果表明:该算法在新场景下的新样本标注准确率高达99.76%.通过迁移学习,新场景的车辆识别分类器性能较通用分类器在检测率和误检率指标上均有显著提升.

Existing classifiers for vehicle recognition are mainly trained offline with a large number of samples,of which the performance may decline dramatically in a new scenario. In order to solve the problem,a sample self-marking transfer learning algorithm for vehicle recognition based on the probabilistic neural network（ PNN） is proposed. First,a sample self-marking strategy is proposed based on multi-cue prior know ledge including complexity,vertical plane and relative velocity.Then,instead of traditional classifiers such as Adaboost,PNN is used to establish the transfer learning model by utilizing its features such as high architecture flexibility,fast training speed and no retraining requirement when new samples are added. Experimental results demonstrate that this algorithm can mark new samples with high accuracy（ 99. 76%）. Besides,new classifier trained in a new scenario with transfer learning performs better in true positive rate and false detection rate than traditional general classifiers.