改进MobileNet-SSD的电梯内电动车识别方法

Improved MobileNet-SSD Recognition Method for Electric Vehicles in Elevators

由于电动车入户充电易造成火灾风险,传统的视频检测方法需要进行数据上传,易造成数据丢失和高延时。针对上述问题,这里提出一种可在嵌入式设备上运行的改进MobileNet-SSD方法,用于识别电梯内电动车目标。在数据预处理阶段,采用CycleGAN方法进行数据增强以提高模型的泛化能力,针对MobileNet网络计算量过大的问题,这里提出引入宽度乘数和分辨率乘数作为超参数以降低模型运算量,经BOHB(Hyperband-Bayesian Optimization)方法对超参数进行优化后获取最优超参数组合。同时为解决权重无法更新和信息丢失问题,用LReLU取代ReLU作为模型的激活函数。实验结果表明,这里的改进MobileNet-SSD算法能在嵌入式设备上快速、准确地识别电动车目标,改进后的MobileNet-SSD模型能将原SSD模型的map提高6%,响应延迟降低20%。

Since the charging of electric vehicles in the home is likely to cause a fire risk,the traditional video detection method re⁃quires data upload,which is likely to cause data loss and high latency.In response to the above problems,this paper proposes an improved MobileNet-SSD method that can be run on embedded devices to identify electric vehicle targets in elevators.In the data preprocessing stage,the CycleGAN method is used for data enhancement to improve the generalization ability of the model.Aim⁃ing at the problem of excessive calculation of the MobileNet network,this paper proposes to introduce the width multiplier and the resolution multiplier as hyperparameters to reduce the amount of model calculations.After optimizing the hyperparameters by the BOHB(Hyperband-Bayesian Optimization)method,the optimal hyperparameter combination is obtained.At the same time,in order to solve the problem of the weight cannot be updated and the information loss,LReLU is used instead of ReLU as the activa⁃tion function of the model.Experimental results show that the improved MobileNet-SSD algorithm in this paper can quickly and accurately identify electric vehicle targets on embedded devices.The improved MobileNet-SSD model can increase the map of the original SSD model by 6%and reduce the response delay by 20%.