Research On Automatic Driving Object Detection Technology Based On CNN

The frequently occurrence of traffic accidents has brought about economic losses and huge personal injuries,with the majority of traffic accidents being caused by human factors.Autonomous driving can achieve mobility without the need for personnel to drive,significantly reducing the incidence of accidents.Visual object detection,as the core technology of autonomous driving perception,has received human attention.Facing different road conditions and weather conditions in actual driving,how to correctly detect multiple types of object on the road is of utmost importance.Based on the above,the main work of this thesis is as follows:A detailed study was conducted on the current object detection algorithms,and finally the YOLO V3 detection algorithm based on convolutional neural networks(CNN)was selected.Based on the actual situation of various weather conditions,road conditions,and multiple types of targets during driving,the BDD100 K dataset was ultimately selected as the training dataset for this article.Add a judgment statement during the label format conversion process: remove labels and files with a label box area of 0,and then use K-means clustering to generate appropriate prior boxes.The DBG module was proposed.Because the positive and negative intervals of Leaky Re LU use different functions to predict the relationship between the positive and negative input values,in view of the number of times activation function is used in the convolutional neural network,the actual application scenarios,and the advantages and disadvantages of the activation function,constructed a DBG module to apply to the non backbone network part.The results indicate that the constructed DBG module has better performance compared to the DBL module in the original YOLO V3 model,m AP increased by 0.011%.The ECA-PANet network was proposed.In response to the poor performance of YOLO V3 in detecting small targets and objects with overlapping sizes,as well as the damage to detail information caused by upsampling,Applying the ECA-PANet network to the YOLO V3 neck section not only allows for more complete integration of underlying and top-level information,ensuring a smaller fusion span,avoiding the impact of large fusion information differences on the expression ability of multi-scale features,but also makes the network pay more attention to "regions of interest",improved the detection accuracy of the detection algorithm.From the perspective of model lightweighting and detection speed,several methods for model lightweighting are analyzed,and finally,hierarchical pruning in model pruning is used to process the network for lightweighting.The autonomous driving detection algorithm constructed in this thesis,EP-YOLO V3-S detection algorithm,has better performance in object detection compared to the original detection algorithm on the dataset in this article and in actual scenarios,and also has excellent detection performance on other datasets.Started from the multiple evaluation indicators of autonomous driving target detection algorithms,this article selects the YOLO V3 object detection algorithm as the basic research framework to construct the EP-YOLO V3-S object detection algorithm.The results indicate that the algorithm proposed in this article can achieve good detection results for multiple types of targets in different driving environments,providing some help and reference for automatic driving target detection problems.