Pedestrian detection method based on feature pyramid network

Abstract

The invention discloses a pedestrian detection method based on a feature pyramid network. The method comprises the following steps of S1, establishing an FOSFPN; S2, reading video frames and performing preprocessing; S3, forwards transmitting frame images, and obtaining detection boxes; S4, performing non-maximum suppression on a detection result to filter repeated and redundant detection boxes; and S5, standardizing the detection boxes exceeding a boundary, thereby obtaining a final detection result. According to the pedestrian detection method based on the feature pyramid network, provided by the invention, the calculation amount can be reduced; pedestrian identification is taken as an application background; and the performance loss caused by network simplification is reduced.

Description

technical field [0001] The invention belongs to the field of image processing and pattern recognition, and relates to a pedestrian detection method based on a feature pyramid network. Background technique [0002] Feature pyramids are a fundamental building block in multi-scale object detection systems. In recent years, deep learning object detection has deliberately avoided pyramid feature representation, because feature pyramids are expensive in terms of computation and memory. Therefore, the author uses the inherent multi-scale and multi-level pyramid structure of the deep convolutional neural network to construct a feature pyramid network. Using a top-down side connection, a high-level semantic feature map is constructed at all scales. This structure is called Feature Pyramid Network (FPN). It has improved significantly in feature extraction. Using FPN on Faster R-CNN, on the COCO dataset, the result exceeds all current single-model detection methods, and can run up to...

AI Technical Summary

Problems solved by technology

[0005] Among them, MTCNN (face detection) adopts small neural network classification and sliding window for fast detection, which achieves good detection effect and speed, but the generalization performance is poor, and it needs to be retrained for specific scenes.

YOLO and SSD use a deep convolutional network to realize the detection process by classifying and biasing the anchor points of each feature map at one time. Such methods are faster in speed and have better generalization performance, but in terms of accuracy loss

The performance of the two-stage Faster RCNN algorithm is the best compared to other types, but due to the intervention of the fully connected layer, the amount of calculation is greatly increased, and it is difficult to apply to industrial application scenarios that require real-time performance.

[0006] On the one hand, the deepening of network complexity will improve the effect of the network, and at the same time increase the amount of calculation

Images