基于多源损失自适应的交通指示灯识别

Traffic light recognition based on adaptive multi⁃source loss

为了提高交通指示灯信号的识别精度,提出一种基于多源损失自适应的交通指示灯识别方法。该方法采用BI⁃LSTM多层自编码对交通指示灯信号进行特征提取,整合后的特征向量作为新的输入,将数据传输至MLP神经网络,再经过softmax层实现数据样本的分类计算,最后采用梯度下降方法,通过模型训练实现模型参数和自适应参数的优化。与一般深度学习单一损失来源不同,该模型具有三个损失来源,分别是编解码损失、对比损失以及交叉熵损失,模型的总损失是由这三个损失以相应的权重叠加而来,权重参数ζ和β是自适应参数,随着模型的训练,ζ和β进行独立学习,最终达到理想值。结果表明多源损失自适应策略对模型自我优化的有效性,提高了模型识别精度。

A traffic light recognition method based on adaptive multi⁃source loss is proposed to improve the recognition accuracy of traffic lights.BI⁃LSTM multi⁃layer auto encoding is used to extract the feature of traffic lights,and the integrated feature vector is used as a new input to transmit the data to MLP neural network.The data samples are classified and calculated by softmax layer.Gradient descent method is used to optimize the model parameters and adaptive parameters by model training.Different from the single⁃source loss of general deep learning,this model has three loss sources,which are encoding and decoding loss,contrast loss and cross entropy loss.The total loss of the model is the superposition of the three losses with corresponding weights.The weight parametersζandβare adaptive parameters.With the training of the model,ζandβlearn independently and finally reach the ideal value.The results show that the strategy of adaptive multi⁃source loss is effective for the self⁃optimization of the model and effectively improves the recognition accuracy of model.