A network method for object detection based on camera projection model

Abstract

The invention discloses an object detection network method based on a camera projection model, including: inputting an image, designing and calculating an anchor frame of the input image; backbone network: inputting the image into the backbone network, and outputting it after passing through multiple feature layers; designing a Razor module ; first encode the ground truth, and then predict; negative sample screening; training and selection of samples, design loss function, training, to obtain the training model; the use of the model; when the training is completed, use the obtained function parameters to deduce the model, The obtained estimate of each anchor box is obtained to obtain the probability of the existence of the target under the anchor box, and the real position in the actual image is obtained by inverse deduction. The article detection network method proposed by the present invention greatly reduces the amount of computation while still retaining the performance of accurate detection in the network. In autonomous driving and surveillance, two industries that rely heavily on camera projection models, it has proven its high efficiency and achieved very good results.

Description

technical field [0001] The present invention relates to the field of network object detection, and more specifically, to an object detection network method and system based on a camera projection model. Background technique [0002] Convolutional neural network (CNN), as a very popular carrier for image recognition and detection, has achieved great success. Based on this technology, many application networks have been derived, such as VGG, ResNet, DenseNet, Yolo and other OneStage networks. . However, these end-to-end (End to End) networks for recognition have many practical problems: First, these networks require a huge amount of computation, making it impossible to actually implement them. The inability to implement includes two levels. One is that the huge amount of calculation leads to higher GPU usage, which increases the deployment cost. The second level is that the huge amount of calculation makes it difficult to achieve real-time calculation. In order to achieve rea...

AI Technical Summary

Problems solved by technology

However, these end-to-end (End to End) networks for recognition have many practical problems: First, these networks require a huge amount of calculations, making it impossible to actually land

The inability to implement includes two levels. One is that the huge amount of calculation leads to higher GPU usage, which increases the deployment cost. The second level is that the huge amount of calculation makes it difficult to achieve real-time calculation. In order to achieve real-time calculation, often need Deploying more computing equipment will cause waste of resources while increasing costs

Second, directly using the simplified network models of these classic networks cannot achieve the performance of the network very well.

Third, some networks with good performance do not design the network for the projection model of the camera, but prefer the network design based on the image itself. This design is more general, but not efficient.

Images