针对特定情况的现有对象检测算法 [英] Existing object detection algorithms for specific case

问题描述

我正在寻找一种算法（在opencv中最好的情况）来检测和跟踪特定低分辨率距离图像中的对象。目标是有一个边界框，最好地描述对象。对于我的用例，我只有16个感兴趣的区域。每个RoI给出了该区域中对象的距离以及该对象占用的空间量（百分比）。以下是信息的外观：



示例



由于每个RoI只获得一个距离，因此只能使用此信息无法区分此RoI中的两个对象。但也许有一种算法可以使占用信息受益。哪些算法可用于检测和跟踪视野内的多个物体和给定的信息。



附加信息：

-The边界框的像素空间固定大小为500,500（输出网格的大小）

-算法因此必须采用4x4像素并在子像素空间中操作才能找到边界框

- 对象不能相互依赖，但可以像示例中那样彼此相邻（第2行，第0列）



我尝试了什么：



我看过几个算法，比如opencv中的Blob-Detector。但它们都使用正常像素值而不是我给出的值，我不确定如何重新定义图像信息，以便它可以与这些正常的计算机视觉算法一起使用。

I'm looking for a algorithm (best case in opencv) to detect and track objects in a specific low resolution distance image. The goal is to have a bounding box, which best describes the object. For my use case I only have 16 Regions of Interest. Each RoI gives me the distance of the object in that Region and the amount of space occupied by that object (percent). Here is how the information could look like:

example

Since I only get one distance per RoI there is no way to differentiate between 2 objects inside this RoI by just using this information. But maybe there is an algorithm that can make benefit of the occupancy information. Which algorithms could be used to detect and track multiple objects inside the field of view with the given information.

additional information:
-The pixelspace for the bounding box is at a fixed size of 500,500 (size of outputgrid)
-The algorithm would therefore have to take the 4x4 pixels and operate in subpixelspace to find the bounding boxes
-Objects can not lie on of of each other, but can be beside each other like in the example (row 2, col 0)

What I have tried:

I've looked at several algorithms like the Blob-Detector in opencv. But they all work with normal pixel values instead of the values that I am given and I'm not sure how I could redefine the picture information, so that it can be used with these normal computer vision algorithms.

推荐答案

看起来你应该在这里使用一些形态图像处理。但是没有方法可以给你一个独特的答案，因为可能有几个blob可能在彼此之上。

Looks like you should use some morphological image processing here. But no method is going to give you a unique answer, as potentially several blobs could be on top of each other.

Tracking意味着运动;因此，可以通过比较一段时间内的帧来检测运动，从而检测物体。

"Tracking" implies "movement"; movement, and therefore "objects", can be detected by comparing frames over time.

这篇关于针对特定情况的现有对象检测算法的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！