Non-rigid multiscale object detection method based on convolutional neural network

Abstract

The invention discloses a non-rigid multiscale object detection method based on a convolutional neural network, which comprises the steps of firstly designing a deep network, performing deep network learning and training for target detection by infrared pedestrian data sets with certain representativeness cvc-09, cvc-14 and self marked small target pedestrian data, wherein the infrared pedestriandata sets have small targets with the pixel being about 20*30 and big target pedestrians at a close range; then designing different detection models based on a hardware platform, and evaluating the detection models. The designed network applies 7*7 grid partition and applies six candidate boxes in allusion to each grid, target area nomination is accomplished, and pedestrian target detection is achieved through classification recognition and location regression analysis. The neural network is applied in the design, and the feature extraction ability of the detection model is increased. The real-time performance of detection is realized based on the idea of grid partition, the number of the candidate boxes is increased by applying data sets with different target scales and multiscale training, and finally the small target detecting ability is enhanced.

Description

technical field [0001] The invention relates to the technical field of real-time object detection, in particular to a non-rigid multi-scale object detection method based on a convolutional neural network. Background technique [0002] Such as image 3 As shown, the neuron is a mathematical model that imitates the biological neural structure. Different inputs (such as different pixels) are summed by weighting (such as the convolution operation in the convolutional neural network), and the nonlinear function ( activation function) to obtain the output of the neural node. Such as Figure 4 As shown, many neural nodes can form a neural layer, and many neural layers can build a neural network. The neural network exhibits a strong fitting ability (learning ability) to the training data due to the input through multi-layer nonlinear transformation (nesting of nonlinear functions), thus showing good predictive ability. [0003] For neural network models, there are two phases, tra...

AI Technical Summary

Problems solved by technology

[0005] The technical problem to be solved by the present invention is: in the real-time detection of targets on mobile devices and small-volume devices, due to the non-rigid multi-scale target shapes and angles in different scenes are changeable, the universality and robustness of traditional detection methods are not high

Images