折腾笔记[10]-使用rust进行ORB角点检测

摘要

打包ORB算法到bye_orb_rs库,使用rust进行ORB角点检测.
Package the ORB algorithm into the bye_orb_rs library, and use Rust for ORB corner detection.

关键词

rust;ORB;FAST;slam;

关键信息

项目地址:[https://github.com/ByeIO/slambook2.rs]

[package]
name = "exp65-rust-ziglang-slambook2"
version = "0.1.0"
edition = "2021"[dependencies]
env_logger = { version = "0.11.6", default-features = false, features = ["auto-color","humantime",
] }# 随机数
rand = "0.8.5"
rand_distr = "0.4.3"
fastrand = "2.3.0"# 线性代数
nalgebra = { version = "0.33.2",features = ["rand"]}
ndarray = "0.16.1"# winit
wgpu = "23.0.1"
winit = "0.30.8"# egui
eframe = "0.30.0"
egui = { version = "0.30.0", features = ["default"
]}
egui_extras = {version = "0.30.0",features = ["default", "image"]}# three_d
three-d = {path = "./static/three-d" , features=["egui-gui"] }
three-d-asset = {version = "0.9",features = ["hdr", "http"] }# sophus
sophus = { version = "0.11.0" }
sophus_autodiff = { version = "0.11.0" }
sophus_geo = "0.11.0"
sophus_image = "0.11.0"
sophus_lie = "0.11.0"
sophus_opt = "0.11.0"
sophus_renderer = "0.11.0"
sophus_sensor = "0.11.0"
sophus_sim = "0.11.0"
sophus_spline = "0.11.0"
sophus_tensor = "0.11.0"
sophus_timeseries = "0.11.0"
sophus_viewer = "0.11.0"
tokio = "1.43.0"
approx = "0.5.1"
bytemuck = "1.21.0"
thingbuf = "0.1.6"# rust-cv计算机视觉
cv = { version = "0.6.0" , features = ["default"] }
cv-core = "0.15.0"
cv-geom = "0.7.0"
cv-pinhole = "0.6.0"
akaze = "0.7.0"
eight-point = "0.8.0"
lambda-twist = "0.7.0"
image = "0.25.5"
imageproc = "0.25.0"# 最小二乘优化
gomez = "0.5.0"# 图优化
factrs = "0.2.0"# ORB角点检测
bye_orb_rs = { path = "./static/bye_orb_rs" }# 依赖覆盖
[patch.crates-io]
pulp = { path = "./static/pulp" }

原理简介

ORB角点检测概述

• rust 角点检测库[https://github.com/ByeIO/bye.orb.rs]
• Rublee E, Rabaud V, Konolige K, et al. ORB: An efficient alternative to SIFT or SURF[C]//2011 International conference on computer vision. Ieee, 2011: 2564-2571.
  关键词:BRIEF,FAST

ORB（Oriented FAST and Rotated BRIEF）是一种高效的角点检测和特征描述算法，由Ethan Rublee等人在2011年提出。ORB结合了FAST角点检测器和BRIEF描述子，并在其基础上进行了改进，使其在保持较高性能的同时，具有更好的旋转不变性和计算效率。

1. FAST角点检测

FAST（Features from Accelerated Segment Test）是一种高效的角点检测算法。其核心思想是通过比较像素点与其周围邻域像素的灰度值来判断是否为角点。FAST算法的步骤如下：

• 选择一个像素点 ( p )，其灰度值为 ( I_p )。
• 设定一个阈值 ( T )。
• 检查以 ( p ) 为中心的圆形邻域（通常为16个像素）内的像素。
• 如果邻域内有连续 ( N ) 个像素的灰度值都大于 ( I_p + T ) 或小于 ( I_p - T )，则 ( p ) 被认为是角点。

FAST算法的优点是计算速度快，适合实时应用。

2. BRIEF描述子

BRIEF（Binary Robust Independent Elementary Features）是一种二进制特征描述子，用于描述图像中的关键点。BRIEF通过比较关键点周围像素对的灰度值来生成一个二进制字符串，作为该关键点的描述子。BRIEF的优点是计算简单、存储空间小，适合大规模图像匹配。

3. ORB的改进

ORB在FAST和BRIEF的基础上进行了以下改进：

• 方向性（Oriented）：ORB通过计算关键点的方向（通常使用灰度质心法），使得BRIEF描述子具有旋转不变性。
• 尺度不变性：ORB通过构建图像金字塔来实现尺度不变性。
• 高效性：ORB在保持较高性能的同时，计算速度比SIFT和SURF更快，适合实时应用。

python的matplotlib设置中文字体

查找系统字体:

brew install font-noto-sans-cjk
fc-list :lang=zh | grep "得意黑"

设置:

from matplotlib import rcParams# 设置中文字体
rcParams['font.sans-serif'] = ['Smiley Sans']  # 指定默认字体为得意黑
# rcParams['font.sans-serif'] = ['Noto Sans CJK SC']  # 指定默认字体为Noto Sans
rcParams['axes.unicode_minus'] = False  # 解决保存图像时负号'-'显示为方块的问题

实现

原始cpp代码:[https://github.com/gaoxiang12/slambook2/blob/master/ch7/orb_cv.cpp]

#include <iostream>
#include <opencv2/core/core.hpp>
#include <opencv2/features2d/features2d.hpp>
#include <opencv2/highgui/highgui.hpp>
#include <chrono>using namespace std;
using namespace cv;int main(int argc, char **argv) {if (argc != 3) {cout << "usage: feature_extraction img1 img2" << endl;return 1;}//-- 读取图像Mat img_1 = imread(argv[1], CV_LOAD_IMAGE_COLOR);Mat img_2 = imread(argv[2], CV_LOAD_IMAGE_COLOR);assert(img_1.data != nullptr && img_2.data != nullptr);//-- 初始化std::vector<KeyPoint> keypoints_1, keypoints_2;Mat descriptors_1, descriptors_2;Ptr<FeatureDetector> detector = ORB::create();Ptr<DescriptorExtractor> descriptor = ORB::create();Ptr<DescriptorMatcher> matcher = DescriptorMatcher::create("BruteForce-Hamming");//-- 第一步:检测 Oriented FAST 角点位置chrono::steady_clock::time_point t1 = chrono::steady_clock::now();detector->detect(img_1, keypoints_1);detector->detect(img_2, keypoints_2);//-- 第二步:根据角点位置计算 BRIEF 描述子descriptor->compute(img_1, keypoints_1, descriptors_1);descriptor->compute(img_2, keypoints_2, descriptors_2);chrono::steady_clock::time_point t2 = chrono::steady_clock::now();chrono::duration<double> time_used = chrono::duration_cast<chrono::duration<double>>(t2 - t1);cout << "extract ORB cost = " << time_used.count() << " seconds. " << endl;Mat outimg1;drawKeypoints(img_1, keypoints_1, outimg1, Scalar::all(-1), DrawMatchesFlags::DEFAULT);imshow("ORB features", outimg1);//-- 第三步:对两幅图像中的BRIEF描述子进行匹配，使用 Hamming 距离vector<DMatch> matches;t1 = chrono::steady_clock::now();matcher->match(descriptors_1, descriptors_2, matches);t2 = chrono::steady_clock::now();time_used = chrono::duration_cast<chrono::duration<double>>(t2 - t1);cout << "match ORB cost = " << time_used.count() << " seconds. " << endl;//-- 第四步:匹配点对筛选// 计算最小距离和最大距离auto min_max = minmax_element(matches.begin(), matches.end(),[](const DMatch &m1, const DMatch &m2) { return m1.distance < m2.distance; });double min_dist = min_max.first->distance;double max_dist = min_max.second->distance;printf("-- Max dist : %f \n", max_dist);printf("-- Min dist : %f \n", min_dist);//当描述子之间的距离大于两倍的最小距离时,即认为匹配有误.但有时候最小距离会非常小,设置一个经验值30作为下限.std::vector<DMatch> good_matches;for (int i = 0; i < descriptors_1.rows; i++) {if (matches[i].distance <= max(2 * min_dist, 30.0)) {good_matches.push_back(matches[i]);}}//-- 第五步:绘制匹配结果Mat img_match;Mat img_goodmatch;drawMatches(img_1, keypoints_1, img_2, keypoints_2, matches, img_match);drawMatches(img_1, keypoints_1, img_2, keypoints_2, good_matches, img_goodmatch);imshow("all matches", img_match);imshow("good matches", img_goodmatch);waitKey(0);return 0;
}

重构的rust代码:

#![allow(dead_code)]
#![allow(unused_variables)]
#![allow(unused_imports)]
#![allow(unused_mut)]
#![allow(unused_assignments)]
#![cfg_attr(not(debug_assertions), windows_subsystem = "windows")]
#![allow(rustdoc::missing_crate_level_docs)]
#![allow(unsafe_code)]
#![allow(clippy::undocumented_unsafe_blocks)]
#![allow(unused_must_use)]
#![allow(non_snake_case)]use image::{open, ImageBuffer, Rgb, DynamicImage
};
use imageproc::{drawing::draw_cross_mut, drawing::draw_line_segment_mut
};use bye_orb_rs::{orb, fast, common::Matchable
};use std::time::Instant;fn main() {let img1_path = "./assets/ch7-1.png";let img2_path = "./assets/ch7-2.png";main_orb(img1_path, img2_path);
}fn main_orb(img1_path: &str, img2_path: &str) {// 读取图像let mut img1 = open(img1_path).unwrap();let mut img2 = open(img2_path).unwrap();// 设置关键点数量let n_keypoints = 500;// 第一步: 检测Oriented FAST角点位置并计算BRIEF描述子let start_time = Instant::now();let img1_keypoints = orb::orb(&mut img1, n_keypoints).unwrap();let img2_keypoints = orb::orb(&mut img2, n_keypoints).unwrap();let end_time = Instant::now();println!("提取ORB特征点耗时: {:?} 秒", end_time - start_time);// 第二步: 使用Hamming距离进行匹配let start_time = Instant::now();let pair_indices = orb::match_brief(&img1_keypoints, &img2_keypoints);let end_time = Instant::now();println!("匹配ORB特征点耗时: {:?} 秒", end_time - start_time);// 第三步: 匹配点对筛选let mut matches: Vec<(usize, usize, f32)> = pair_indices.iter().map(|&(i, j)| (i, j, img1_keypoints[i].distance(&img2_keypoints[j]) as f32)).collect();matches.sort_by(|a, b| a.2.partial_cmp(&b.2).unwrap());let min_dist = matches[0].2;let max_dist = matches[matches.len() - 1].2;println!("-- 最大距离: {}", max_dist);println!("-- 最小距离: {}", min_dist);// 当描述子之间的距离大于两倍的最小距离时，认为匹配有误。但最小距离可能会非常小，设置一个经验值30作为下限。let good_matches: Vec<(usize, usize, f32)> = matches.iter().filter(|&&(_, _, dist)| dist <= (2.0 * min_dist).max(30.0)).cloned().collect();// 第四步: 绘制匹配结果let mut img1_rgb = img1.to_rgb8();let mut img2_rgb = img2.to_rgb8();// 绘制所有匹配点for &(i, j, _) in &matches {let kp1 = (img1_keypoints[i].x, img1_keypoints[i].y);let kp2 = (img2_keypoints[j].x, img2_keypoints[j].y);let color = Rgb([0, 255, 0]);draw_cross_mut(&mut img1_rgb, color, kp1.0 as i32, kp1.1 as i32);draw_cross_mut(&mut img2_rgb, color, kp2.0 as i32, kp2.1 as i32);}// 保存绘制结果img1_rgb.save("all_matches.png").unwrap();img2_rgb.save("all_matches2.png").unwrap();// 绘制筛选后的匹配点let mut img1_rgb_good = img1.to_rgb8();let mut img2_rgb_good = img2.to_rgb8();for &(i, j, _) in &good_matches {let kp1 = (img1_keypoints[i].x, img1_keypoints[i].y);let kp2 = (img2_keypoints[j].x, img2_keypoints[j].y);let color = Rgb([0, 255, 0]);draw_cross_mut(&mut img1_rgb_good, color, kp1.0 as i32, kp1.1 as i32);draw_cross_mut(&mut img2_rgb_good, color, kp2.0 as i32, kp2.1 as i32);}// 保存绘制结果img1_rgb_good.save("good_matches.png").unwrap();img2_rgb_good.save("good_matches2.png").unwrap();// 第五步: 创建并排放置的图像let (width1, height1) = img1_rgb.dimensions();let (width2, height2) = img2_rgb.dimensions();let total_width = width1 + width2;let max_height = height1.max(height2);let mut combined_image = ImageBuffer::new(total_width, max_height);// 将第一张图片复制到组合图像的左侧for y in 0..height1 {for x in 0..width1 {combined_image.put_pixel(x, y, *img1_rgb.get_pixel(x, y));}}// 将第二张图片复制到组合图像的右侧for y in 0..height2 {for x in 0..width2 {combined_image.put_pixel(width1 + x, y, *img2_rgb.get_pixel(x, y));}}// 绘制连接线for &(i, j, _) in &good_matches {let kp1 = (img1_keypoints[i].x, img1_keypoints[i].y);let kp2 = (img2_keypoints[j].x, img2_keypoints[j].y);let color = Rgb([0, 255, 0]);draw_line_segment_mut(&mut combined_image,(kp1.0 as f32, kp1.1 as f32),((width1 as f32 + kp2.0 as f32), kp2.1 as f32),color,);}// 保存并排放置的图像combined_image.save("combined_matches.png").unwrap();
}

python代码:

import cv2
import numpy as np
import matplotlib.pyplot as plt
import time
from matplotlib import rcParams# 设置中文字体
rcParams['font.sans-serif'] = ['Smiley Sans']  # 指定默认字体为得意黑
rcParams['axes.unicode_minus'] = False  # 解决保存图像时负号'-'显示为方块的问题def main(img1_path, img2_path):# 读取图像img1 = cv2.imread(img1_path, cv2.IMREAD_COLOR)img2 = cv2.imread(img2_path, cv2.IMREAD_COLOR)assert img1 is not None and img2 is not None, "图片读取失败"# 初始化ORB检测器orb = cv2.ORB_create()# 第一步: 检测Oriented FAST角点位置start_time = time.time()keypoints1 = orb.detect(img1, None)keypoints2 = orb.detect(img2, None)# 第二步: 计算BRIEF描述子keypoints1, descriptors1 = orb.compute(img1, keypoints1)keypoints2, descriptors2 = orb.compute(img2, keypoints2)end_time = time.time()print(f"提取ORB特征点耗时: {end_time - start_time} 秒")# 绘制特征点outimg1 = cv2.drawKeypoints(img1, keypoints1, None, color=(0, 255, 0), flags=0)plt.imshow(outimg1)plt.title("ORB特征点")plt.show()# 第三步: 使用Hamming距离进行匹配bf = cv2.BFMatcher(cv2.NORM_HAMMING, crossCheck=True)start_time = time.time()matches = bf.match(descriptors1, descriptors2)end_time = time.time()print(f"匹配ORB特征点耗时: {end_time - start_time} 秒")# 第四步: 匹配点对筛选matches = sorted(matches, key=lambda x: x.distance)min_dist = matches[0].distancemax_dist = matches[-1].distanceprint(f"-- 最大距离: {max_dist}")print(f"-- 最小距离: {min_dist}")# 当描述子之间的距离大于两倍的最小距离时，认为匹配有误。但最小距离可能会非常小，设置一个经验值30作为下限。good_matches = [m for m in matches if m.distance <= max(2 * min_dist, 30.0)]# 第五步: 绘制匹配结果img_matches = cv2.drawMatches(img1, keypoints1, img2, keypoints2, matches, None, flags=2)img_good_matches = cv2.drawMatches(img1, keypoints1, img2, keypoints2, good_matches, None, flags=2)plt.imshow(img_matches)plt.title("所有匹配点")plt.show()plt.imshow(img_good_matches)plt.title("筛选后的匹配点")plt.show()if __name__ == "__main__":import sysmain("../assets/ch7-1.png", "../assets/ch7-2.png")

效果

rust输出:

提取ORB特征点耗时: 1.921584042s 秒
匹配ORB特征点耗时: 1.893269416s 秒
-- 最大距离: 49
-- 最小距离: 0

python输出:

提取ORB特征点耗时: 0.055989980697631836 秒
匹配ORB特征点耗时: 0.001383066177368164 秒
-- 最大距离: 81.0
-- 最小距离: 4.0

所有检测结果1	所有检测结果2	最佳检测结果1	最佳检测结果2	直线连接筛选后的匹配点	python结果1	python结果2