Research On Road Traffic Behavior Recognition Based On Deep Learning

Driving environment perception has always been an extremely important part of the automatic driving system.Efficient and accurate identification of the surrounding environment and the state and behavior information of traffic participants is of great significance for the decision-making and control of intelligent driving assistance.In recent years,methods and technologies for driving environment perception have emerged in an endless stream,and traditional methods have gradually been replaced by deep learning methods with high growth potential and strong learnability.In the case of low popularity of self-vehicle intelligence and edge computing capabilities,algorithms based on deep learning still have problems such as cumbersome training and deployment,and high computing overhead.In response to above problems,this thesis takes road traffic participants as the main research object,and builds a deep neural network model for object detection,multi-object tracking and behavior recognition based on deep learning methods,focusing on improving network performance and focusing on the lightweight design of the model.This thesis has mainly done the following work:(1)This thesis constructs an object detection network based on the feature extraction of the re-parameterized structure,detects objects such as pedestrians and vehicles,optimizes the network structure design,decouples the training process and the deployment process.In the case of ensuring high model recognition accuracy,reduce the computing resource overhead of model training and model application.(2)Based on the detection-based tracking algorithm in the object tracking task,use the object detection algorithm designed above and the attention-enhanced object tracking algorithm to further optimize the performance of the object tracking algorithm,enhance the object tracking effect in camera motion scenes and improve the robustness of the algorithm.(3)In order to solve the problem of pedestrian target positioning and long-term behavior feature information extraction in traffic scenes,the thesis integrates the aforementioned object detection network and object tracking network,which effectively improves the behavior recognition ability in multi-target scenarios.The magnitude-scale backbone network design and time-sliding feature extraction capabilities can effectively reduce the computational complexity of the model,making it easier to deploy the model and perform targeted fast calculations.