Research On Driver's Hand-held Call Status Detection Method Based On Deep Learning

The driver's hand-held calling behavior while driving will seriously distract his attention,making the risk of road safety accidents more than four times higher than during normal driving.At present,domestic and foreign scholars have conducted relatively few researches on the detection of the specific behavior status of the driver's hand-held call,mainly focusing on non-visual and visual-based detection research.With the rapid rise of deep learning in recent years,its automatic learning features,high accuracy and strong robustness based on convolutional neural networks have brought new light to the fields of image classification,target detection and target tracking.In this thesis,the target detection method based on deep learning is applied to the research of driver's hand-held call status recognition in the field of transportation,and the focus is on the target detection method based on regional proposals.The main research work is as follows:Firstly,construct a data set of the driver's hand-held call status.This article collects real-time video data from the Bittar transportation operation platform,converts these video data into continuous video frames,and amplifies them by mirroring to create the data set required for the experiment.Before the experiment,the data set is unified into the VOC2007 data format and divided into the training set and the test set.Then,the LabelImg image annotation tool is used to classify and label the training data set to form a label file for subsequent training and testing of the network.Secondly,an improved mobile object detection method of Faster R-CNN is proposed.Aiming at the influence of a large amount of background noise and light and dark changes on the image,the image is pre-processed simply by data enhancement.Introduce the idea of expanding convolution,improve the residual structure and integrate it into the detection sub-network part,to alleviate the problem that the input image gradually becomes smaller after feature extraction.The improved Faster R-CNN detection model is trained using a four-step crossover method and the corresponding multitask loss function is set.Finally,the performance of the model before and after improvement is compared through experiments.Experiments show that the accuracy of the improved detection model has reached 91.42%,compared with the original model image detection accuracy has been effectively improved.Finally,a method for detecting the status of the driver's hand-held conversation based on strategy fusion is proposed.Based on the improved network structure,different optimization strategies are merged to reduce the cases of missed detection and false detection.Among them,design an appropriate anchor frame strategy and adopt multi-scale training to reduce the leakage detection;then introduce the OHEM(Online Hard Example Mining)strategy into the network model,and train the negative negative samples repeatedly to enhance the network to the negative negative samples.The ability to identify,thereby reducing false detections.Finally,through multiple sets of comparative experiments,hyperparameters(dropout value,batch size and confidence threshold)suitable for the data set in this thesis are selected to further optimize the detection performance.Experiments show that the performance of the detection model after strategy fusion has been further improved,and the performance has been improved by 3.26% while ensuring real-time.The paper has 29 pictures,12 tables,and 82 references.