Skip to content

O
openpose

sq975 / openpose
与 Fork 源项目一致

从无法访问的项目Fork

通知 7
Star 0
Fork 0
O
openpose

前往新版Gitcode，体验更适合开发者的 AI 搜索 >>

提交 d9e745a0 编写于 作者: F fragalfernando 提交者： Gines
浏览文件
操作

Greedy LK matching + AVX/SSE4 optimizations (#597) 

- Greedy/Matrix version of LucasKanade vs OpenPose matcher in identification
- AVX and SSE4 optimizations flags.
上级 825dfcae
隐藏空白更改
内联 并排
Showing with 186 addition and 36 deletion
CMakeLists.txt
浏览文件 @ d9e745a0
......@@ -59,8 +59,23 @@ endif (WIN32)
# C++ additional flags
if (CMAKE_COMPILER_IS_GNUCXX)
# Select the Enhanced Instruction Set
set(INSTRUCTION_SET NONE CACHE STRING "Enable Enhanced Instruction Set")
set_property(CACHE INSTRUCTION_SET PROPERTY STRINGS NONE SSE4 AVX)
if (${INSTRUCTION_SET} MATCHES "SSE4")
add_definitions("-DWITH_SSE4")
set(SIMD_FLAGS "${SIMD_FLAGS} -msse4.1")
endif(${INSTRUCTION_SET} MATCHES "SSE4")
if (${INSTRUCTION_SET} MATCHES "AVX")
add_definitions("-DWITH_AVX")
set(SIMD_FLAGS "${SIMD_FLAGS} -mavx")
endif(${INSTRUCTION_SET} MATCHES "AVX")
message(STATUS "GCC detected, adding compile flags")
set(OP_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fopenmp -Wpedantic -Wall -Wextra -Wfatal-errors")
set(OP_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${SIMD_FLAGS} -fopenmp -Wpedantic -Wall -Wextra -Wfatal-errors")
endif (CMAKE_COMPILER_IS_GNUCXX)
......
include/openpose/experimental/tracking/personIdExtractor.hpp
浏览文件 @ d9e745a0
......@@ -2,7 +2,9 @@
#define OPENPOSE_TRACKING_PERSON_ID_EXTRACTOR_HPP
#include <atomic>
#include <tuple>
#include <unordered_map>
#include <unordered_set>
#include <openpose/core/common.hpp>
namespace op
......
src/openpose/experimental/tracking/personIdExtractor.cpp
浏览文件 @ d9e745a0
......@@ -108,9 +108,9 @@ namespace op
}
void initializeLK(std::unordered_map<int,PersonEntry>& personEntries,
long long& mNextPersonId,
const Array<float>& poseKeypoints,
const float confidenceThreshold)
long long& mNextPersonId,
const Array<float>& poseKeypoints,
const float confidenceThreshold)
{
try
{
......@@ -142,36 +142,53 @@ namespace op
}
}
Array<long long> matchLKAndOP(std::unordered_map<int,PersonEntry>& personEntries,
long long& nextPersonId,
const std::vector<PersonEntry>& openposePersonEntries,
const cv::Mat& imagePrevious,
const float inlierRatioThreshold,
const float distanceThreshold)
bool compareCandidates(std::tuple<float, int, int> a, std::tuple<float, int, int> b)
{
return std::get<0>(a) > std::get<0>(b);
}
Array<long long> matchLKAndOPGreedy(std::unordered_map<int,PersonEntry>& personEntries,
long long& nextPersonId,
const std::vector<PersonEntry>& openposePersonEntries,
const cv::Mat& imagePrevious,
const float inlierRatioThreshold,
const float distanceThreshold)
{
try
{
Array<long long> poseIds{(int)openposePersonEntries.size(), -1};
std::unordered_map<int, PersonEntry> pendingQueue;
std::vector<bool> processed((int)openposePersonEntries.size(), false);
std::unordered_set<int> used;
bool converged = false;
if (!openposePersonEntries.empty())
while (!openposePersonEntries.empty() && !converged)
{
const auto numberKeypoints = openposePersonEntries[0].keypoints.size();
std::vector<std::tuple<float, int, int>> candidates;
float bestScore = 0.0f;
converged = true;
for (auto i = 0u; i < openposePersonEntries.size(); i++)
{
auto& poseId = poseIds.at(i);
if (poseIds.at(i) != -1)
continue;
const auto& openposePersonEntry = openposePersonEntries.at(i);
const auto personDistanceThreshold = fastMax(10.f,
distanceThreshold*float(std::sqrt(imagePrevious.cols*imagePrevious.rows)) / 960.f);
// Find best correspondance in the LK set
auto bestMatch = -1ll;
auto bestScore = 0.f;
for (const auto& personEntry : personEntries)
{
if (used.find(personEntry.first) != used.end())
continue;
const auto& element = personEntry.second;
auto inliers = 0;
auto active = 0;
auto distance = 0.f;
auto total_distance = 0.0f;
// Security checks
if (element.status.size() != numberKeypoints)
......@@ -184,6 +201,7 @@ namespace op
if (openposePersonEntry.keypoints.size() != numberKeypoints)
error("openposePersonEntry.keypoints.size() != numberKeypoints",
__LINE__, __FUNCTION__, __FILE__);
// Iterate through all keypoints
for (auto kp = 0u; kp < numberKeypoints; kp++)
{
......@@ -191,8 +209,10 @@ namespace op
if (!element.status[kp] && !openposePersonEntry.status[kp])
{
active++;
const auto distance = getEuclideanDistance(element.keypoints[kp],
openposePersonEntry.keypoints[kp]);
distance = getEuclideanDistance(element.keypoints[kp],
openposePersonEntry.keypoints[kp]);
total_distance += distance;
if (distance < personDistanceThreshold)
inliers++;
}
......@@ -201,26 +221,46 @@ namespace op
if (active > 0)
{
const auto score = inliers / (float)active;
if (score > bestScore && score >= inlierRatioThreshold)
if (score == bestScore && score >= inlierRatioThreshold)
{
candidates.push_back(std::make_tuple(total_distance,i, personEntry.first));
bestScore = score;
}
else if (score > bestScore && score >= inlierRatioThreshold)
{
bestScore = score;
bestMatch = personEntry.first;
candidates.clear();
candidates.push_back(std::make_tuple(total_distance, i, personEntry.first));
}
}
}
// Found a best match, update LK table and poseIds
if (bestMatch != -1)
poseId = bestMatch;
else
poseId = nextPersonId++;
}
std::sort(candidates.begin(), candidates.end(), compareCandidates);
while (candidates.size())
{
auto top_candidate = candidates.back();
candidates.pop_back();
auto idx_lk = std::get<2>(top_candidate);
auto idx_op = std::get<1>(top_candidate);
if (used.find(idx_lk) != used.end())
continue;
poseIds[idx_op] = idx_lk;
used.insert(idx_lk);
converged = false;
pendingQueue[poseId] = openposePersonEntry;
}
}
// Update LK table with pending queue
for (auto& pendingQueueEntry: pendingQueue)
personEntries[pendingQueueEntry.first] = pendingQueueEntry.second;
for (auto i = 0u; i < openposePersonEntries.size(); i++)
{
if (poseIds[i] == -1)
poseIds[i] = nextPersonId++;
const auto& openposePersonEntry = openposePersonEntries.at(i);
personEntries[poseIds[i]] = openposePersonEntry;
}
return poseIds;
}
......@@ -231,6 +271,97 @@ namespace op
}
}
// Array<long long> matchLKAndOP(std::unordered_map<int,PersonEntry>& personEntries,
// long long& nextPersonId,
// const std::vector<PersonEntry>& openposePersonEntries,
// const cv::Mat& imagePrevious,
// const float inlierRatioThreshold,
// const float distanceThreshold)
// {
// try
// {
// Array<long long> poseIds{(int)openposePersonEntries.size(), -1};
// std::unordered_map<int, PersonEntry> pendingQueue;
// if (!openposePersonEntries.empty())
// {
// const auto numberKeypoints = openposePersonEntries[0].keypoints.size();
// for (auto i = 0u; i < openposePersonEntries.size(); i++)
// {
// auto& poseId = poseIds.at(i);
// const auto& openposePersonEntry = openposePersonEntries.at(i);
// const auto personDistanceThreshold = fastMax(10.f,
// distanceThreshold*float(std::sqrt(imagePrevious.cols*imagePrevious.rows)) / 960.f);
// // Find best correspondance in the LK set
// auto bestMatch = -1ll;
// auto bestScore = 0.f;
// for (const auto& personEntry : personEntries)
// {
// const auto& element = personEntry.second;
// auto inliers = 0;
// auto active = 0;
// // Security checks
// if (element.status.size() != numberKeypoints)
// error("element.status.size() != numberKeypoints ||", __LINE__, __FUNCTION__, __FILE__);
// if (openposePersonEntry.status.size() != numberKeypoints)
// error("openposePersonEntry.status.size() != numberKeypoints",
// __LINE__, __FUNCTION__, __FILE__);
// if (element.keypoints.size() != numberKeypoints)
// error("element.keypoints.size() != numberKeypoints ||", __LINE__, __FUNCTION__, __FILE__);
// if (openposePersonEntry.keypoints.size() != numberKeypoints)
// error("openposePersonEntry.keypoints.size() != numberKeypoints",
// __LINE__, __FUNCTION__, __FILE__);
// // Iterate through all keypoints
// for (auto kp = 0u; kp < numberKeypoints; kp++)
// {
// // If enough threshold
// if (!element.status[kp] && !openposePersonEntry.status[kp])
// {
// active++;
// const auto distance = getEuclideanDistance(element.keypoints[kp],
// openposePersonEntry.keypoints[kp]);
// if (distance < personDistanceThreshold)
// inliers++;
// }
// }
// if (active > 0)
// {
// const auto score = inliers / (float)active;
// if (score > bestScore && score >= inlierRatioThreshold)
// {
// bestScore = score;
// bestMatch = personEntry.first;
// }
// }
// }
// // Found a best match, update LK table and poseIds
// if (bestMatch != -1)
// poseId = bestMatch;
// else
// poseId = nextPersonId++;
// pendingQueue[poseId] = openposePersonEntry;
// }
// }
// // Update LK table with pending queue
// for (auto& pendingQueueEntry: pendingQueue)
// personEntries[pendingQueueEntry.first] = pendingQueueEntry.second;
// return poseIds;
// }
// catch (const std::exception& e)
// {
// error(e.what(), __LINE__, __FUNCTION__, __FILE__);
// return Array<long long>{};
// }
// }
PersonIdExtractor::PersonIdExtractor(const float confidenceThreshold, const float inlierRatioThreshold,
const float distanceThreshold, const int numberFramesToDeletePerson) :
mConfidenceThreshold{confidenceThreshold},
......@@ -290,8 +421,10 @@ namespace op
}
// Get poseIds and update LKset according to OpenPose set
poseIds = matchLKAndOP(mPersonEntries, mNextPersonId, openposePersonEntries, mImagePrevious,
mInlierRatioThreshold, mDistanceThreshold);
// poseIds = matchLKAndOP(
poseIds = matchLKAndOPGreedy(
mPersonEntries, mNextPersonId, openposePersonEntries, mImagePrevious, mInlierRatioThreshold,
mDistanceThreshold);
return poseIds;
}
......
src/openpose/experimental/tracking/pyramidalLK.cpp
浏览文件 @ d9e745a0
......@@ -5,12 +5,12 @@
#include <openpose/utilities/profiler.hpp>
#include <openpose/experimental/tracking/pyramidalLK.hpp>
#if defined ( __SSE4_1__)
#if defined (WITH_SSE4)
#include <emmintrin.h>
#include "smmintrin.h"
#endif
#if defined (__AVX__)
#if defined (WITH_AVX)
#include <immintrin.h>
#endif
......@@ -32,7 +32,7 @@
namespace op
{
#if defined ( __SSE4_1__)
#if defined (WITH_SSE4)
float sse_dot_product(std::vector<float> &av, std::vector<float> &bv)
{
......@@ -68,7 +68,7 @@ namespace op
}
#endif
#if defined (__AVX__)
#if defined (WITH_AVX)
float avx_dot_product(std::vector<float> &av, std::vector<float> &bv)
{
......@@ -114,13 +114,13 @@ namespace op
auto sumYT = 0.f;
auto sumXY = 0.f;
#if defined (__AVX__)
#if defined (WITH_AVX)
sumXX = avx_dot_product(ix,ix);
sumYY = avx_dot_product(iy,iy);
sumXY = avx_dot_product(ix,iy);
sumXT = avx_dot_product(ix,it);
sumYT = avx_dot_product(iy,it);
#elif defined ( __SSE4_1__)
#elif defined (WITH_SSE4)
sumXX = sse_dot_product(ix,ix);
sumYY = sse_dot_product(iy,iy);
sumXY = sse_dot_product(ix,iy);
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或