Vehicle detection method based on multi-branch cyclic self-attention network and cyclic frame regression

Abstract

The invention discloses a vehicle detection method based on a multi-branch cyclic self-attention network and cyclic frame regression. The vehicle detection method comprises the following steps: constructing a trunk network for vehicle detection; predicting a candidate box by using an RPN network, extracting an instance feature map according to the candidate box, circularly predicting the self-attention map and calculating a new instance feature map to obtain a final instance feature map; circularly selecting a feature map output by the basic network according to the predicted detection frame by using circular frame regression; and expanding the computing network to multiple branches by using a multi-branch network structure, and fusing a multi-branch detection result. Compared with a traditional vehicle detection method based on deep learning, the vehicle detection method based on the multi-branch cyclic self-attention network and the cyclic frame regression has the advantages that more accurate features of the vehicle can be obtained, the confidence coefficient of a detection frame of the vehicle is improved, a more accurate detection frame can be obtained, and the detection accuracy is improved.

Description

technical field [0001] The invention relates to the technical field of vehicle detection, in particular to a vehicle detection method based on a multi-branch cyclic self-attention network and cyclic frame regression. Background technique [0002] Vehicle detection is an important part of driver assistance system (ADAS) and automatic driving system (ADS). A vehicle detection algorithm with higher accuracy is of great significance to the safety of automatic driving systems and assisted driving systems. Due to the strong generalization ability and fitting ability of deep learning, the vehicle detection algorithm based on deep learning has greatly improved in terms of accuracy. At present, the vehicle detection algorithms based on deep learning mainly include Fast RCNN, Faster RCNN, SSD, etc. In the target detection algorithm based on deep learning, in the training phase, the input image is input to the convolutional neural network to extract features, and the matching algorit...

AI Technical Summary

Problems solved by technology

In addition, traditional deep learning algorithms only select features once or select features twice as features for subsequent fine-tuning.

The above two shortcomings lead to the fact that the features selected by the current vehicle detection algorithm are still not accurate enough, the predicted detection frame is not accurate enough, and the accuracy rate is not high

Therefore, it is urgent to provide a vehicle detection method based on multi-branch cyclic self-attention network and cyclic border regression to solve the problem of low detection accuracy of occluded vehicles in traditional deep learning vehicle detection algorithms.

Images