| With the rapid development of the automobile industry,the number of cars is increasing,and the large-scale increase of motor vehicle base increases the incidence of road traffic accidents to a certain extent.According to the investigation,the primary reason for the increase in accident rate is that drivers carry out operations unrelated to driving tasks in the process of driving,so it is particularly important to detect related violations.Based on the analysis of behavior change characteristics,this thesis proposes a dangerous driving behavior detection algorithm based on supervised learning combined with deep learning technology,which is connected with embedded devices to accurately identify violations and quickly warn,which is of great significance to the construction of intelligent traffic.The main research contents are as follows:（1）Aiming at the problem of unbalanced detection accuracy and delay performance of dangerous driving behavior,an improved filter redundancy box SSD（F-SSD）algorithm based on Filter redundancy box modules was proposed.F-SSD algorithm firstly uses Resnet-50 to extract more source image information and uses end-to-end processing to increase model fit.Then,a pyramid-like structure is constructed,which takes into account the feature information of the bottom layer and the top layer.Single-linear interpolation method is used for up-sampling to improve the robustness of multi-scale objects and enhance the three-dimensional space-time features.Finally,the filter redundancy box module was designed,and the Gaussian kernel function was introduced to select the self-adaptive threshold,so as to optimize the positive and negative sample allocation strategy in the training process and reduce the computational cost in the network iteration.Through the experimental comparison of the algorithm before and after the improvement,the improved F-SSD algorithm is improved in accuracy and meets the real-time requirements.（2）Due to the insufficient performance of real-time video stream due to the large amount of data,the method of randomly picking frames per unit time is adopted to improve the detection effect,but it is easy to lose key information.To solve this problem,an ORB based video key frame extraction method is adopted in this thesis.By calculating the similarity degree of feature descriptors of frame images,the most informative frame is extracted as the key frame to retain important characterization information.In order to improve the efficiency of video retrieval,when F-SSD algorithm is loaded to identify violations,important frame nodes are synchronically recorded,and the violation clips are intercepted and retained to realize intelligent processing of video information.（3）A dangerous driving behavior warning system based on edge computing is designed.The hardware detection part of the system adopts NVIDIA Jetson AGX Xavier development kit.By loading the dangerous driving behavior detection model,the hardware equipment acquisition and identification function is completed.The software platform adopts B/S separation architecture with intelligent data interaction capability to manage relevant data,which mainly includes functional modules such as real-time monitoring,driving behavior recording and big data analysis.If detected irregularities,edge end equipment according to the violation time quantitative risk level,risk level too high voice warning prompt,the background and the related illegal data to the software,monitors can remote monitoring in the management of end,traceability illegal detailed information,and through the interception of fragment quick check,after confirm there are irregularities,take corresponding measures,we will establish a rigorous and sound supervision system.Figure [47] table [8] reference [71]... |
PDF Full Text Request