Research On Traffic Sign Detection Algorithm Based On Convolutional Neural Network

Traffic sign detection is an important part of the assisted driving system.In recent years,the application of convolutional neural networks has made breakthroughs in object detection algorithm.The existing object detection algorithm based on convolutional neural network has achieved good performance on the general dataset,but when it is directly applied to the traffic sign detection task,it is difficult to obtain better detection results.This thesis proposes to study the traffic sign detection algorithm based on convolutional neural network,and makes innovative research results in the following aspects:Firstly,it is proposed to improve the sampling strategy of the SSD algorithm by using the Intersection over Ground-truth standard.Compared with the general object detection scenario,the object in traffic sign detection only occupies a very small proportion of the image.The SSD algorithm makes it difficult to effectively sample the object using the Intersection over Union based sampling strategy,and compresses the object size during training.The problem is that the training sample has a single scale and the training scale is incomplete.In this thesis,the Intersection over Ground-truth standard is used to control the completeness of the sampling object,and the object multi-scale sampling is achieved by randomly adjusting the sampling frame size,so that the detection accuracy of the algorithm is significantly improved.Secondly,a traffic sign detection algorithm based on step-by-step deep learning is proposed.The algorithm firstly meshes the image evenly,and then constructs a lightweight full convolution network to quickly predict the mesh containing the object and determine the potential area of the traffic sign.Then,the adjacent object candidate regions are combined according to the object adhesion criterion to obtain a final object candidate region.Since the object size range is greatly reduced at this time,we design a streamlined deep learning network to quickly identify candidate object regions to obtain the final object detection results.The algorithm achieved 91% accuracy and 92% recall on the TT100 K dataset,and achieved a detection speed of 20.9FPS on TITAN X.Both the detection accuracy and the real-time performance of the algorithm are much better than the traditional algorithm.Thirdly,it is proposed to optimize and accelerate the traffic sign detection algorithm based on step-by-step deep learning by using depth separable convolution.On mobile platforms,conventional algorithms are difficult to detect traffic signs quickly and accurately.This thesis proposes the optimization and acceleration of the object grid prediction network and the traffic sign detection network using the depth separable convolution to the step-bystep detection algorithm,which significantly improves the detection speed of the algorithm.The accelerated algorithm sacrificed 6.6% of the detection accuracy,but the detection speed increased by 23 times,reached 6FPS on the mobile platform,and obtained 89.5% accuracy and 86.7% recall rate for the medium and large scale objects on the TT100 K dataset.