Anti-occlusion pedestrian detection method based on attention mechanism

Abstract

The invention discloses an anti-occlusion pedestrian detection method based on an attention mechanism, and the method comprises the steps: obtaining the multilayer convolution features of a pedestrianimage, and generating a new feature layer containing semantic information and detail information through the structural fusion of a new feature pyramid; adopting k-means clustering to obtain the sizeof a preset pedestrian bounding box; predicting regression bias and classification confidence of all pedestrian bounding boxes; calculating a cross entropy loss function and an improved adaptive occlusion perception regression loss function to obtain an overall loss function; carrying out iterative optimization and continuous training to obtain a detection model; and sending the to-be-detected picture into the detection model, and removing a frame with a high overlapping degree by adopting a non-maximum suppression method to finally obtain a detection result. The method can improve the detection capability of shielded pedestrians, and has higher detection precision and robustness.

Description

technical field [0001] The invention belongs to the fields of computer vision, pattern recognition, deep learning and the like, and in particular relates to a method for detecting the position of pedestrians in complex traffic scene images or video frames. Background technique [0002] With the rise of the artificial intelligence boom and the advent of the era of big data, computer vision technology based on images and videos to understand the world has been vigorously developed. As an important branch of general target detection, pedestrian detection has important applications in intelligent transportation, automatic driving, video surveillance, intelligent robots and other fields. In addition, pedestrian detection algorithms with high detection accuracy and good robustness are the premise and foundation of many advanced tasks in computer vision, such as pedestrian pose recognition, behavior analysis, multi-target tracking, and pedestrian re-identification. Therefore, how ...

AI Technical Summary

Problems solved by technology

The above-mentioned representative method of pedestrian detection based on convolutional neural network has a sharp decline in detection performance under severe occlusion. The main reason is that occlusion will lead to the loss of local features of pedestrians, and at the same time introduce interference information that does not belong to pedestrians, and the neural network cannot autonomously distinguish different features. In addition, the loss function of common detection algorithms does not consider the influence of occlusion degree on the position candidate box, especially in the case of very dense pedestrians, a large number of occlusion candidate boxes will mislead the update of the gradient, and ultimately affect the performance of detection

Images