Research On Key Technologies Of Multi-target Perception For Complex Road Scenes

As the quality of people’s lives continuously improves,the demand for transportation tools increases year by year,making road safety issues increasingly prominent.Vehicle-assisted driving systems can collect surrounding environmental information,predict and forcibly intervene in potential hazards,thereby reducing the occurrence of traffic accidents.Environmental perception,as the primary link of the vehicle-assisted driving system,has a direct impact on the performance of the entire system.In recent years,deep learning-based object detection algorithms have made significant progress in the field of image processing and have been widely applied to environmental perception to improve detection effects.However,complex and changing road traffic scenarios,such as small and occluded objects,pose challenges to object detection algorithms.Simultaneously,in-vehicle mobile devices pose certain limitations concerning model complexity and computational efficiency.Therefore,enhancing the detection accuracy of object detection models in complex road scenarios and the detection efficiency within constrained in-vehicle hardware resources holds significant research implications.In this thesis,a series of improvements are made to the mainstream object detection algorithm YOLOv5 s,designing the high-performance YOLOv5s-MISC object detection algorithm.To improve the model’s adaptability to domestic road traffic environments,this thesis also builds a complex domestic road scenario dataset to train the model.To tackle the issue of insufficient detection accuracy for occluded and small objects in road scenes,we enhance YOLOv5 s by integrating a cascaded Swin Transformer Block to build the CSwin module,applied within the model.By using parameters to control model depth,we accommodate different feature fusion scales,thereby augmenting the model’s global feature fusion capabilities and ameliorating occluded object detection.Simultaneously,we devise a novel YOLOv5s-MS feature fusion model,introducing shallow features from the backbone network into the model.This approach allows for the extraction and sharing of shallow features to preserve small object location details to the greatest extent possible.Road traffic object detection is constrained by the computational resources of in-vehicle mobile devices,necessitating high model detection efficiency.While ensuring the improved model’s performance,we apply a novel Involution operator to design the Slim C3 module,which is integrated into the feature fusion network.This module extracts channel-wise information and eliminates redundancy generated by convolution operations,thereby reducing model parameters and enhancing computational efficiency.To further improve the model’s training effectiveness on the dataset,the loss function is improved to address low-overlap sample dominance in gradient updates and dataset sample imbalance during model training.A penalty factor is set in the original CIOU regression loss function to weaken the weight of low-overlap samples,thus mitigating the model training problems caused by lowoverlap samples.Additionally,the classification loss and confidence loss are modified to Focal Loss loss function to enhance the model’s performance on imbalanced datasets.Through comprehensive validation of the proposed method on both a custom-built dataset and the public KITTI dataset,the experimental results demonstrate that: the model incorporating the CSwin module achieved an increase of 1.7% and 1.2% in m AP@0.5 on the custom-built dataset and KITTI dataset,respectively;the YOLOv5 sMS feature fusion model showed a 4.9% improvement in m AP@0.5:0.95 for small objects and a 5.3% increase in recall rate;the YOLOv5s-Inv Neck model,which introduced the Involution operator,reduced the number of parameters by 9.3% without affecting accuracy;enhancements to the loss function effectively mitigated the impact of low-overlap samples and simultaneously improved the model’s performance on imbalanced data.Ultimately,the combined application of these improvements to the YOLOv5s-MISC model resulted in a 3% overall increase in m AP@0.5,effectively enhancing the detection of occluded and small objects while maintaining a lightweight structure.This demonstrates that the methods proposed in this study can effectively address the demand for high detection accuracy in computational resource-limited scenarios.