Driving Behavior Recognition Based On Lightweight Convolutional Neural Network

In the existing research on driving behavior recognition,most methods need to extract specific features in advances,such as head posture angle,gaze direction,electroencephalogram,hand,and body joint positions.These features are not always easy to obtain,and some even require specific hardware equipment,which will add additional time or financial costs.At the same time,the quality of feature selection and the limitations of feature extraction methods further reduce the accuracy of these methods.Deep learning methods provide an efficient and practical solution for driving behavior recognition research.Existing deep learning methods face the following three technical problems: insufficient sample richness of datasets,backbone convolutional neural networks cannot be applied to resource-constrained on-board computing devices,and existing lightweight convolutional neural networks sacrifice accuracy Performance is traded for a decrease in parameters and calculations.To solve the above problems,this research starts with the sample richness of the dataset and the design of a lightweight convolutional nerve dedicated to driving behavior,and designs and creates a self-acquired driving behavior image dataset to solve the problem of serious lack of sample richness in the existing dataset,And for the task of driving behavior recognition,the idea of frequency decomposition is applied to the design of the lightweight convolutional neural network,and a lightweight convolutional neural network designed for driving behavior recognition task is created.details as follows:1.In order to resolve the problem of insufficient diversity of public dataset samples in the field of driving behavior recognition,the largest distracted driving image dataset in the field of driving behavior research has been created,which is called Lilong Driving Behavior(LDB).The LDB dataset is based on a standard collection process and made regarding the public data set State Farm.It collected six distracted driving behaviors of2,468 people.Compared with the State Farm data set,the sample richness of this data set is more than 20 times that of it,which effectively improves the robustness of the method.2.In order to resolve the problem that the heavyweight backbone network cannot be applied to in-vehicle embedded devices,an Octave-like Convolution Mixed(OLCM)building block was created.The OLCM building block uses the frequency decomposition idea of eight-dimensional convolution to decompose the feature map of the network into the two branches of high and low frequency through point convolution,to extract the high and low-frequency features respectively;then use the information fusion with a lightweight attention mechanism to merge and merge Compress high and low-frequency information.The OLCM building block reduces connection density and spatial redundancy and improves the efficiency and performance of multi-scale feature extraction.3.In view of the huge amount of parameters and calculations of the final classifier of the fully connected structure,an efficient final stage of automatic feature selection is proposed.A fully convolutional structure is used to replace the traditional heavy-duty fully connected classifier,and the squeeze and excitation operations are used to automatically select important features to improve classification performance while reducing the amount of calculation and parameters.The experimental results show that,compared with the existing methods,the driving behavior recognition based on the lightweight Octave-like Convolution Mixed Neural Network(OLCMNet)achieves the best performance on the vehicle embedded platform.Also,the network trained on the LDB dataset has achieved better accuracy performance.For the Statefarm public dataset,the accuracy of OLCMNet is 89.53%,and the latency is32.8ms.For the LDB dataset,the accuracy of OLCMNet is 95.98%,and the latency is32.8ms.At the same time,the richness of the sample is very important to the practical application of the method.In actual in-vehicle online experiments,OLCMNet trained on the LDB dataset obtained an average accuracy of 97%,while the average accuracy of OLCMNet trained on the State Farm dataset was only 81.8%.