| Object detection in traffic scenes has important applications in autonomous driving and intelligent traffic management.However,the detection effect in actual applications can be affected by unfavorable factors such as target occlusion,bad weather and low light conditions.In the field of deep learning,current research mostly focuses on proposing better learnable methods to enhance the expression ability of features.The performance of these methods depends heavily on the parameters learned by the model on the training set,which leads to serious degradation of detection performance when these methods are transferred to test scenarios with large differences from the training set.To address this challenge,this study proposes an iterative attention model based on traditional mixed probability models.The attention algorithm introduces the mixed student t-distribution into the implementation process of the attention algorithm,and uses its excellent outlier fitting ability to perform effective high-dimensional semantic clustering and optimization of localization features learned by convolutional neural networks.The attention model is implemented in an iterative manner similar to EM algorithm,and its core part does not require neural network participation in learning,thus having non-learning characteristics.Based on this attention model,this study further proposes an object detection framework that combines learnable networks with non-learning attention.This framework uses conventional deep learning networks to extract features and initially locate objects,and uses the above-mentioned iterative attention algorithm to optimize localization features.To fully utilize the computational results of the attention model,this study proposes to use corresponding features in the semantic activation classification branch perceived by the classification in the obtained attention map.Finally,to ensure that the attention algorithm can accurately discover areas of interest under extreme conditions,this study proposes an optional foreground mask module for learning and activating foreground areas in input features of attention models.Experiments on multiple difficult traffic datasets show that the proposed detector has excellent robustness and generalization in various difficult scenarios.At the same time,the proposed iterative attention mechanism has good versatility and can be quickly transferred to other visual tasks like semantic segmentation. |
