A Study Of Driver Action Recognition In A Realtime Environment

The current popular driver action recognition models have some problems:(1)attention crash problem,(2)cannot be used well with a small amount of fine-tuning in multiple models,it often requires retraining new datasets,adding extra workload to the task.(3)Some existing generic models have large number of parameters,low training and inference efficiency,and are not friendly for personal computers with small computational power.In this paper,we propose a dynamic adaptive driver action recognition model to solve the above problem,which can directly migrate models from other domains to the driver action recognition task with a small amount of fine-tuning,and the number of model parameters and inference speed are also improved compared with other traditional algorithms.The overall algorithm implementation process is as follows:(1)driver action video preprocessing.(2)input the data set into a model consisting of residual joins performed by the above modules for training.(3)randomly select a sample or a set of samples not in the original video for testing with common criteria such as accuracy,model parameters,and model inference speed.Its main architecture is divided into the following parts:(1)Predictor module,through the prediction of the input video frame sequence,to derive the patches to be discarded,so as to reduce the model parameters,increase the training and inference speed,and reduce the computer arithmetic pressure.(2)Re-attention module,the dataset produced in this paper has the characteristics of high action repetition rate and large similarity,which can cause the attention of the Transformer model used to collapse,thus making the accuracy rate no longer increase as the model reaches a certain depth after recognition accuracy,after analysis and experimentation,the re-attention module can be a good solution to the above problem After analysis and experiments,the re-attention module can solve the above problem.(3)Adaptor module,which is a plug-and-play fine-tuning module,its main role is to reduce the difficulty of migrating models from other domains to the driver action recognition task.Previously,if models from other domains and tasks were to be applied to a specific recognition job,they often needed to be trained from scratch.This makes the research more difficult and takes longer to complete the task,and if the amount of parameters of the original model is too large this requires an unacceptable amount of overhead.This module can fine-tune as few parameters as possible to maintain higher recognition accuracy,providing convenient methods and ideas.Experiments are conducted on several publicly available datasets to increase the persuasiveness and feasibility of the model and algorithm.The experiments prove that the proposed method has the advantages of more stable accuracy,lower model parameters,faster inference,and ease of migration in driver action recognition compared to traditional driver action recognition methods,and is feasible for engineering and research purposes.