Double Exponential Smoothing Filter for Azure Kinect Body Tracking SDK

k4a_double_exponential_filter.h

/*
 k4a_double_exponential_filter.h

 This file contains holt double exponential smoothing filter for filtering joints.
 It was ported for Azure Kinect Body Tracking SDK based on following implementation.
 https://social.msdn.microsoft.com/Forums/en-US/045b058a-ae3a-4d01-beb6-b756631b4b42

 std::unordered_map<int32_t, double_exponential_filter> double_exponential_filter;

 while( true )
 {
     
     for( const k4abt_body_t& body : bodies )
     {
         double_exponential_filter[body.id].update( body, body_frame.get_timestamp() );
         k4abt_body_t filtered_body = double_exponential_filter[body.id].get_result();
         for( const k4abt_joint_t& filtered_joint : filtered_body.skeleton.joints )
         {
             // Access to filtered joint position ...
         }
     }

     std::chrono::microseconds threshold( 50000 );
     std::for_each( double_exponential_filter.begin(), double_exponential_filter.end(),
         [&]( auto& filter ){
             std::chrono::microseconds interval( body_frame.get_timestamp() - filter.second.get_timestamp() );
             if( interval.count() > threshold.count() )
             {
                 double_exponential_filter.erase( filter.first );
             }
         }
     );
 }

 Copyright (c) 2019 Tsukasa Sugiura <t.sugiura0204@gmail.com>
 Licensed under the MIT license.

 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal
 in the Software without restriction, including without limitation the rights
 to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
 copies of the Software, and to permit persons to whom the Software is
 furnished to do so, subject to the following conditions:

 The above copyright notice and this permission notice shall be included in all
 copies or substantial portions of the Software.

 THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
 AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 SOFTWARE.
*/

#ifndef K4A_DOUBLE_EXPONENTIAL_FILTER
#define K4A_DOUBLE_EXPONENTIAL_FILTER

#include <k4a/k4a.h>
#include <k4abt.h>

#include <array>
#include <chrono>

k4a_float3_t operator * ( const float d, const k4a_float3_t& v )
{
    k4a_float3_t r;
    r.xyz.x = d * v.xyz.x;
    r.xyz.y = d * v.xyz.y;
    r.xyz.z = d * v.xyz.z;
    return r;
}

k4a_float3_t operator + ( const k4a_float3_t& v1, const k4a_float3_t& v2 )
{
    k4a_float3_t r;
    r.xyz.x = v1.xyz.x + v2.xyz.x;
    r.xyz.y = v1.xyz.y + v2.xyz.y;
    r.xyz.z = v1.xyz.z + v2.xyz.z;
    return r;
}

k4a_float3_t operator - ( const k4a_float3_t& v1, const k4a_float3_t& v2 )
{
    k4a_float3_t r;
    r.xyz.x = v1.xyz.x - v2.xyz.x;
    r.xyz.y = v1.xyz.y - v2.xyz.y;
    r.xyz.z = v1.xyz.z - v2.xyz.z;
    return r;
}

class double_exponential_filter
{
public:
    double_exponential_filter()
        : smoothing_( 0.5f ),
          correction_( 0.5f ),
          prediction_( 0.5f ),
          jitter_radius_( 300.0f ),
          max_deviation_radius_( 500.0f ),
          id_( 0 ),
          timestamp_( std::chrono::microseconds( 0 ) )
    {
        for( int32_t i = 0; i < K4ABT_JOINT_COUNT; i++ )
        {
            filtered_joints_[i].position = initialize();
            history_[i]                  = double_exponential_data();
        }
    }

    void set_parameter( const float smoothing, const float correction, const float prediction, const float jitter_radius, const float max_deviation_radius )
    {
        smoothing_            = smoothing;
        correction_           = correction;
        prediction_           = prediction;
        jitter_radius_        = jitter_radius;
        max_deviation_radius_ = max_deviation_radius;
    }

    void reset()
    {
        set_parameter( 0.5f, 0.5f, 0.5f, 300.0f, 500.0f );
        id_        = 0;
        timestamp_ = std::chrono::microseconds( 0 );
        for( int32_t i = 0; i < K4ABT_JOINT_COUNT; i++ )
        {
            filtered_joints_[i].position = initialize();
            history_[i] = double_exponential_data();
        }
    }

    void update( const k4abt_body_t& body, const std::chrono::microseconds timestamp )
    {
        id_        = body.id;
        timestamp_ = timestamp;

        for( int32_t i = 0; i < K4ABT_JOINT_COUNT; i++ )
        {
            k4a_float3_t raw_position      = body.skeleton.joints[i].position;
            k4a_float3_t filtered_position = initialize();
            k4a_float3_t trend             = initialize();

            k4a_float3_t previous_raw_position      = history_[i].raw_position;
            k4a_float3_t previous_filtered_position = history_[i].filtered_position;
            k4a_float3_t previous_trend             = history_[i].trend;

            k4a_float3_t diff = initialize();
            float length      = 0.0f;

            if( !is_valid( raw_position ) )
            {
                history_[i].frame_count = 0;
            }

            if( history_[i].frame_count == 0 )
            {
                filtered_position = raw_position;
                trend             = initialize();
                history_[i].frame_count++;
            }
            else if( history_[i].frame_count == 1 )
            {
                filtered_position = 0.5f * ( raw_position + previous_raw_position );
                diff              = filtered_position - previous_filtered_position;
                trend             = ( correction_ * diff ) + ( ( 1.0f - correction_ ) * previous_trend );
                history_[i].frame_count++;
            }
            else
            {
                diff   = raw_position - previous_filtered_position;
                length = std::sqrt( ( diff.xyz.x * diff.xyz.x ) + ( diff.xyz.y * diff.xyz.y ) + ( diff.xyz.z * diff.xyz.z ) );
                length = std::fabs( length );

                if( length <= jitter_radius_ )
                {
                    filtered_position = ( ( length / jitter_radius_ ) * raw_position ) + ( ( 1.0f - length / jitter_radius_ ) * previous_filtered_position );
                }
                else
                {
                    filtered_position = raw_position;
                }

                filtered_position = ( ( 1.0f - smoothing_ ) * filtered_position ) + ( smoothing_ * ( previous_filtered_position + previous_trend ) );

                diff  = filtered_position - previous_filtered_position;
                trend = ( correction_ * diff ) + ( ( 1.0f - correction_ ) * previous_trend );
            }

            k4a_float3_t predicted_position = filtered_position + ( prediction_ * trend );

            diff   = predicted_position - raw_position;
            length = std::sqrt( ( diff.xyz.x * diff.xyz.x ) + ( diff.xyz.y * diff.xyz.y ) + ( diff.xyz.z * diff.xyz.z ) );
            length = std::fabs( length );

            if( length > max_deviation_radius_ )
            {
                predicted_position = ( ( ( max_deviation_radius_ / length ) * predicted_position ) + ( ( 1.0f - max_deviation_radius_ / length ) * raw_position ) );
            }

            history_[i].raw_position      = raw_position;
            history_[i].filtered_position = filtered_position;
            history_[i].trend             = trend;

            filtered_joints_[i].position.xyz.x = filtered_position.xyz.x;
            filtered_joints_[i].position.xyz.y = filtered_position.xyz.y;
            filtered_joints_[i].position.xyz.z = filtered_position.xyz.z;
            filtered_joints_[i].orientation    = body.skeleton.joints[i].orientation;
        }
    }

    const k4abt_body_t get_result()
    {
        k4abt_body_t result;
        result.id = id_;
        for( int32_t i = 0; i < K4ABT_JOINT_COUNT; i++ )
        {
            result.skeleton.joints[i].position.xyz.x = filtered_joints_[i].position.xyz.x;
            result.skeleton.joints[i].position.xyz.y = filtered_joints_[i].position.xyz.y;
            result.skeleton.joints[i].position.xyz.z = filtered_joints_[i].position.xyz.z;
            result.skeleton.joints[i].orientation    = filtered_joints_[i].orientation;
        }
        return result;
    }

    const std::chrono::microseconds get_timestamp()
    {
        return timestamp_;
    }
private:
    inline bool is_valid( k4a_float3_t vector )
    {
        return ( vector.xyz.x != 0.0f ||
                 vector.xyz.y != 0.0f ||
                 vector.xyz.z != 0.0f );
    }

    static k4a_float3_t initialize()
    {
        k4a_float3_t vector;
        vector.xyz.x = 0.0f;
        vector.xyz.x = 0.0f;
        vector.xyz.x = 0.0f;
        return vector;
    }

    struct double_exponential_data
    {
        k4a_float3_t raw_position;
        k4a_float3_t filtered_position;
        k4a_float3_t trend;
        uint32_t     frame_count;

        double_exponential_data()
        {
            raw_position      = initialize();
            filtered_position = initialize();
            trend             = initialize();
            frame_count       = 0;
        }
    };

    std::array<k4abt_joint_t, K4ABT_JOINT_COUNT> filtered_joints_;
    std::array <double_exponential_data, K4ABT_JOINT_COUNT> history_;
    float smoothing_;
    float correction_;
    float prediction_;
    float jitter_radius_;
    float max_deviation_radius_;
    uint32_t id_;
    std::chrono::microseconds timestamp_;
};

#endif

main.cpp

#include <iostream>
#include <k4a/k4a.h>
#include <k4a/k4a.hpp>
#include <opencv2/opencv.hpp>

// Include Converter Utility Header
#include "util.h"

// Include C++ Wrapper for Azure Kinect Body Tracking SDK
#include "k4abt.hpp"

// Include Ｄouble Exponential Smoothing Filter for Azure Kinect Body Tracking SDK
#include "k4a_double_exponential_filter.h"
#include <unordered_map>

int main( int argc, char* argv[] )
{
    try{
        // Get Connected Devices
        const int32_t device_count = k4a::device::get_installed_count();
        if( device_count == 0 ){
            throw k4a::error( "Failed to found device!" );
        }

        // Open Default Device
        k4a::device device = k4a::device::open( K4A_DEVICE_DEFAULT );

        // Start Cameras with Configuration
        k4a_device_configuration_t configuration = K4A_DEVICE_CONFIG_INIT_DISABLE_ALL;
        configuration.color_format             = K4A_IMAGE_FORMAT_COLOR_BGRA32;
        configuration.color_resolution         = K4A_COLOR_RESOLUTION_720P;
        configuration.depth_mode               = K4A_DEPTH_MODE_NFOV_UNBINNED;
        configuration.synchronized_images_only = true;
        device.start_cameras( &configuration );

        // Initialize Transformation
        k4a::calibration calibration = device.get_calibration( configuration.depth_mode, configuration.color_resolution );
        k4a::transformation transformation = k4a::transformation( calibration );

        // Create Tracker
        k4abt::tracker tracker = k4abt::tracker::create( calibration );
        if( !tracker ){
            throw k4a::error( "Failed to create tracker!" );
        }

        // Color Table for Visualize
        std::vector<cv::Vec3b> colors;
        colors.push_back( cv::Vec3b( 255,   0,   0 ) );
        colors.push_back( cv::Vec3b(   0, 255,   0 ) );
        colors.push_back( cv::Vec3b(   0,   0, 255 ) );
        colors.push_back( cv::Vec3b( 255, 255,   0 ) );
        colors.push_back( cv::Vec3b(   0, 255, 255 ) );
        colors.push_back( cv::Vec3b( 255,   0, 255 ) );

        // Initialize Double Exponential Filter
        std::unordered_map<int32_t, double_exponential_filter> double_exponential_filter;

        while( true )
        {
            // Get Capture
            k4a::capture capture;
            const std::chrono::milliseconds timeout( K4A_WAIT_INFINITE );
            const bool result = device.get_capture( &capture, timeout );
            if( !result ){
                break;
            }

            // Get Color Image
            k4a::image color_image = capture.get_color_image();
            cv::Mat color = k4a::get_mat( color_image );

            // Get Depth Image
            k4a::image depth_image = capture.get_depth_image();
            cv::Mat depth = k4a::get_mat( depth_image );

            // Enqueue Capture
            tracker.enqueue_capture( capture );

            // Pop Tracker Result
            k4abt::frame body_frame = tracker.pop_result();

            // Get Body Index Map
            k4a::image body_index_map_image = body_frame.get_body_index_map();
            cv::Mat body_index_map = k4a::get_mat( body_index_map_image );

            // Get Body Skeleton
            std::vector<k4abt_body_t> bodies = body_frame.get_bodies();

            // Clear Handle
            capture.reset();
            color_image.reset();
            depth_image.reset();

            // Scaling Depth
            depth.convertTo( depth, CV_8U, -255.0 / 5000.0, 255.0 );

            // Get Image that used for Inference
            k4a::capture body_capture = body_frame.get_capture();
            k4a::image skeleton_image = body_capture.get_color_image();
            cv::Mat skeleton = k4a::get_mat( skeleton_image );

            // Draw Body Skeleton
            for( k4abt_body_t& body : bodies ){
                // Draw Raw Skeleton
                for( const k4abt_joint_t& joint : body.skeleton.joints ){
                    k4a_float2_t position;
                    const bool result = calibration.convert_3d_to_2d( joint.position, K4A_CALIBRATION_TYPE_DEPTH, K4A_CALIBRATION_TYPE_COLOR, &position );
                    if( !result ){
                        continue;
                    }
                    const int32_t id = body.id;
                    const cv::Point point( static_cast<int32_t>( position.xy.x ), static_cast<int32_t>( position.xy.y ) );
                    cv::circle( skeleton, point, 5, colors[( id - 1 ) % colors.size()], -1 );
                }

                // Smooting Filter
                double_exponential_filter[body.id].update( body, body_frame.get_timestamp() );

                // Draw Filtered Skeleton
                k4abt_body_t filtered_body = double_exponential_filter[body.id].get_result();
                for( const k4abt_joint_t& filtered_joint : filtered_body.skeleton.joints ){
                    k4a_float2_t filtered_position;
                    const bool result = calibration.convert_3d_to_2d( filtered_joint.position, K4A_CALIBRATION_TYPE_DEPTH, K4A_CALIBRATION_TYPE_COLOR, &filtered_position );
                    if( !result ){
                        continue;
                    }
                    const int32_t id = filtered_body.id;
                    const cv::Point filtered_point( static_cast<int32_t>( filtered_position.xy.x ), static_cast<int32_t>( filtered_position.xy.y ) );
                    cv::circle( skeleton, filtered_point, 8, colors[( id - 1 ) % colors.size()], 1 );
                }
            }

            // Clean-Up Unused Filters (Sometimes)
            const std::chrono::microseconds threshold( 50000 );
            std::for_each( double_exponential_filter.begin(), double_exponential_filter.end(),
                [&]( auto& filter ){
                    std::chrono::microseconds interval( body_frame.get_timestamp() - filter.second.get_timestamp() );
                    if( interval.count() > threshold.count() ){
                        double_exponential_filter.erase( filter.first );
                    }
                }
            );

            // Clear Handle
            body_frame.reset();
            body_capture.reset();
            skeleton_image.reset();

            // Show Image
            cv::imshow( "skeleton", skeleton );
            const int32_t key = cv::waitKey( 30 );
            if( key == 'q' ){
                break;
            }
        }

        // Close Tracker
        tracker.destroy();

        // Close Transformation
        transformation.destroy();

        // Close Device
        device.close();

        // Close Window
        cv::destroyAllWindows();
    }
    catch( const k4a::error& error ){
        std::cout << error.what() << std::endl;
    }

    return 0;
}