Research On Traffic Sign Detection Algorithm Based On Deep Learning

Object detection takes advantage of computer technologies to localize and recognize specific targets autonomously in real-world scenarios.This technology is the basis and premise of a large number of advanced computer vision tasks.While traditional object detection has good performance with computational speed,the hand-craft features are lack of robustness for occlusion and deformation of targets in complicated natural environments.Recent advances in object detection are owed to the success of the convolutional neural networks(CNNs).By stacking a series of convolutional layers interleaved with activation function and down-sampling,CNNs are capable of learning complex hierarchical feature representations of images.Traffic sign detection is an inevitable and arduous task in many real-world applications.In particularly,localizing and classifying traffic signs accurately is highly related to the safety in the field of advanced driver assistance systems(ASAD).But traffic signs usually exist in the image as small targets.Existing object detection pipelines usually show superior performance for large objects with high resolution but fail to detect objects with tiny size such as traffic signs.To handle this problem,the exploratory research on the traffic sign detection based on deep learning is carried out.(1)Attention-based neural network for traffic sign detection.Attention mechanism has seen significant interest years as the powerful addition to deep neural networks in last several.This thesis proposes a novel end-to-end architecture that improves the performance of small object detection by combing attention mechanism with Faster R-CNN.Specially,this architecture focuses on channel-wise features and utilizes attention mechanism to enhance the feature responses by explicitly modeling the interdependencies between channel-wise features.Finally,the regression of bounding boxes and the classification of traffic signs are generated after selecting the discriminative features by attention.Extensive evaluations on the largest traffic sign dataset demonstrate that the attention mechanism improves the performance of detecting objects at different scales,especially at the small scale.(2)Reverse connection with attention mechanism network for traffic sign detection.High-level feature map is extracted by pooling spatially coarser,but semantically stronger,and low-level feature map is of lower-level semantics,but its activations are more accurately localized as it was subsampled fewer times.This thesis uses attention mechanism to select features which are then merged via reverse connections.Those feature maps are enhanced to the appropriate resolution for detection.The experimental results show that the proposed framework can obtain higher prediction accuracy compared with the original framework.