Research On Pedestrian And Vehicle Detection Methods For Automotive Driver Assistance Systems

With the increasing number of cars and the increasingly complex traffic environment,the automotive driver assistance systems for driving safety is becoming increasingly prominent.The continuous development of target detection technology has made it possible for the popularization of automotive driver assistance systems.Pedestrian and vehicle detection is the most important detection task in the driving road scenario,and in order to improve its detection performance and thus promote the popularity of automotive driver assistance systems,this thesis proposes pedestrian and vehicle detection methods for automotive driver assistance systems.The main points of this thesis are as follows:Firstly,the technical basis of pedestrian and vehicle detection is described.By comparing and analysing traditional target detection algorithms and deep learning-based target detection algorithms,the deep learning-based target detection algorithm with high detection accuracy,strong learning ability and high robustness is selected for better application in vehicle assisted driving systems.In this thesis,a home-made dataset was created by photographing in the field and collecting on the internet.The data were amplified using mirroring,light and dark adjustment,and the introduction of Gaussian noise,and then annotate it and add a small part of UC Berkeley dataset to form the final dataset.The dataset provides good support for model analysis and experimental comparison,and makes up for the deficiency that most of the current public datasets are foreign road scenes,making the research more suitable for the domestic environment.The modelling study of pedestrian and vehicle detection was carried out.Firstly,two twostage algorithms,Fast R-CNN and Faster R-CNN,and two single-stage algorithms,SSD and YOLOv5,were theoretically analysed,and then the four algorithms were modelled based on the dataset of this thesis.The detection accuracy and detection speed of each model were compared through the analysis of experimental results,and finally the YOLOv5 model with stronger performance was selected as the benchmark model.A study on the improvement of the YOLOv5 model was carried out.By proposing the CBAMC3 module,using the bi-directional feature pyramid module,improving the spatial pyramid pooling module,improving the bounding box loss function,and improving the postprocessing method.Based on the dataset of this thesis,experiments are designed to compare the improved model horizontally with the original model,ablation experiments to verify the effectiveness and specific performance improvement of each improvement scheme,and vertically with other classical models and two representative improved YOLOv5 models.The data analysis proves the better comprehensive performance of the improved models in this thesis.The results show that the YOLOv5 model is selected as the benchmark model through theoretical and experimental analysis of typical models.The improved YOLOv5 model and the experimental analysis prove that the comprehensive performance of the improved model in this thesis is improved,and the better pedestrian and vehicle detection effect is achieved,which can meet the pedestrian and vehicle detection tasks in the actual driving road scenarios.