Research On Traffic Sign Detection Method Oriented To Unmanned Visual Perception

Driverless technology plays a positive role in reducing the volume of traffic accidents and ensuring the safety of road network operations.Environmental perception is one of the key methods for highly intelligent autonomous driving.The detection and recognition of driving scenarios by vehicles on the road is achieved through on-board sensors and perception algorithms,which are used as the basis for making corresponding behavioral decisions,thus realizing the autonomous driving capability without any intervention.As the driverless level evolves,the requirements for vision perception algorithms are also increasing.Although the existing object detection algorithm has excellent performance,it cannot be deployed on on-board hardware due to its high computational complexity and cannot meet the immediate needs of driverless vehicles for detection capability and efficiency.Moreover,the centralized model training process can lead to the leakage of individual data and has high performance requirements on the servers that perform model training and data storage.In response to these problems,this paper takes the driverless environment perception task as the background and the traffic sign detection task as the entry point to conduct the following research:1.For the visual perception task,the typical one-stage object detection algorithm is analyzed from the structure,and the existing optimization methods as well as the current state of research are compared and summarized in a targeted manner,so as to establish the research direction of the environment perception object detection model for this article.2.Considering the requirements of the environment perception task on the algorithm performance,an improved one-stage lightweight perception model EG-YOLOv5 is proposed.Inspired by the bottleneck structure and the attention mechanism,the inverted residual structure ESGBlock incorporating the attention mechanism is proposed to address the problems of high-dimensional feature information loss and high model computational complexity,and a lightweight backbone network is constructed based on it.Compared with the baseline model,the model requires only 2.77 M parameters to achieve a real-time inference speed of 35.6FPS on the CPU,effectively improving the problem that traditional detection models are difficult to employ in practice.3.Aiming at the problem that the traffic sign detection task in visual perception is subject to many disturbing factors from natural scenes and the degradation of detection accuracy due to the reduced amount of model parameters,a response-based objectness scaled knowledge distillation method with temperature coefficients is used to train the lightweight model.The predictive knowledge of the high-performance model is transferred to the lightweight perception model by supervised training to achieve performance compensation of the object detection results.Traffic signs are manually augmented and Slicing Aided Hyper Inference techniques are introduced in the sign image preprocessing stage to effectively slice large resolution images to achieve augmentation of sample quantities and more retention of semantic information.Experimental results on two challenging traffic sign datasets show that the distillation model of the method can achieve an effective balance between detection precision and inference speed,and outperforms the baseline model for most categories and scales.4.In response to the problem that the training of traditional deep learning models requires the collection of massive data,and personal information is disseminated and used without restriction,a method based on federated learning of unmanned visual perception is proposed.The travel data with location information is stored by individual users and local model training is performed without data exchange.Encrypted training parameters of multiple vehicles are aggregated based on wireless communication and other technologies and returned by a trusted server for the next round of model training process until the model converges.Compared with existing methods,the security and privacy of vehicle data are significantly improved,and personalized updates of in-vehicle perception models can be achieved under limited computational conditions.