Video-Stabilization/VideoStab.m at main · cuixing158/Video-Stabilization · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
classdef VideoStab < handle
    properties
        smoothedMat
        k
        errscaleX
        errscaleY
        errthetha
        errtransX
        errtransY

        Q_scaleX
        Q_scaleY
        Q_thetha
        Q_transX
        Q_transY

        R_scaleX
        R_scaleY
        R_thetha
        R_transX
        R_transY

        sum_scaleX
        sum_scaleY
        sum_thetha
        sum_transX
        sum_transY

        scaleX
        scaleY
        thetha
        transX
        transY

        HORIZONTAL_BORDER_CROP = 20
    end

    methods
        function obj = VideoStab()
            % Kalman filter parameters
            Q1 = 0.004;
            R1 = 0.5;

            obj.smoothedMat = zeros(2, 3);
            obj.k = 1;

            obj.errscaleX = 1;
            obj.errscaleY = 1;
            obj.errthetha = 1;
            obj.errtransX = 1;
            obj.errtransY = 1;

            obj.Q_scaleX = Q1;
            obj.Q_scaleY = Q1;
            obj.Q_thetha = Q1;
            obj.Q_transX = Q1;
            obj.Q_transY = Q1;

            obj.R_scaleX = R1;
            obj.R_scaleY = R1;
            obj.R_thetha = R1;
            obj.R_transX = R1;
            obj.R_transY = R1;

            obj.sum_scaleX = 0;
            obj.sum_scaleY = 0;
            obj.sum_thetha = 0;
            obj.sum_transX = 0;
            obj.sum_transY = 0;

            obj.scaleX = 0;
            obj.scaleY = 0;
            obj.thetha = 0;
            obj.transX = 0;
            obj.transY = 0;
        end

        function smoothedFrame = stabilize(obj, frame_1, frame_2)
            % 转灰度
            frame1 = im2gray(frame_1);
            frame2 = im2gray(frame_2);

            vert_border = round(obj.HORIZONTAL_BORDER_CROP * size(frame_1,1) / size(frame_1,2));

            % 特征点检测 + 光流跟踪
            points1 = detectMinEigenFeatures(frame1,'MinQuality',0.01,'FilterSize',7);
            features1 = points1.Location;

            tracker = vision.PointTracker('MaxBidirectionalError', 3);
            initialize(tracker, features1, frame1);
            [features2, valid] = step(tracker, frame2);

            goodFeatures1 = features1(valid,:);
            goodFeatures2 = features2(valid,:);

            % 估计 4 自由度仿射变换，scale,theta,tx,ty
            tform = estgeotform2d(goodFeatures1,goodFeatures2,"similarity");
            affine = tform.A(1:2,:);

            dx = affine(1,3);
            dy = affine(2,3);
            da = atan2(affine(2,1), affine(1,1)); % 弧度
            ds_x = affine(1,1) / cos(da);
            ds_y = affine(2,2) / cos(da);

            sx = ds_x;
            sy = ds_y;

            obj.sum_transX = obj.sum_transX + dx;
            obj.sum_transY = obj.sum_transY + dy;
            obj.sum_thetha = obj.sum_thetha + da;
            obj.sum_scaleX = obj.sum_scaleX + ds_x;
            obj.sum_scaleY = obj.sum_scaleY + ds_y;

            % 第一次不做预测
            if obj.k == 1
                obj.k = obj.k + 1;
            else
                obj.Kalman_Filter();
            end

            % Compute differences between smoothed and raw accumulated parameters.
            diff_scaleX = obj.scaleX - obj.sum_scaleX;
            diff_scaleY = obj.scaleY - obj.sum_scaleY;
            diff_transX = obj.transX - obj.sum_transX;
            diff_transY = obj.transY - obj.sum_transY;
            diff_thetha = obj.thetha - obj.sum_thetha;

            ds_x = ds_x + diff_scaleX;
            ds_y = ds_y + diff_scaleY;
            dx = dx + diff_transX;
            dy = dy + diff_transY;
            da = da + diff_thetha;


            % 平滑后的仿射矩阵
            obj.smoothedMat(1,1) = sx * cos(da);
            obj.smoothedMat(1,2) = sx * -sin(da);
            obj.smoothedMat(2,1) = sy * sin(da);
            obj.smoothedMat(2,2) = sy * cos(da);
            obj.smoothedMat(1,3) = dx;
            obj.smoothedMat(2,3) = dy;

            % 相似变换
            tformSmooth = simtform2d([obj.smoothedMat; 0 0 1]);
            smoothedFrame = imwarp(frame_1, tformSmooth, 'OutputView', imref2d(size(frame_2)));

            % 裁剪黑边
            smoothedFrame = smoothedFrame(vert_border+1:end-vert_border, ...
                                          obj.HORIZONTAL_BORDER_CROP+1:end-obj.HORIZONTAL_BORDER_CROP,:);
        end

        function Kalman_Filter(obj)
            frame_1_scaleX = obj.scaleX;
            frame_1_scaleY = obj.scaleY;
            frame_1_thetha = obj.thetha;
            frame_1_transX = obj.transX;
            frame_1_transY = obj.transY;

            % 预测误差协方差: Add process noise Q (simple prediction step without motion model).
            frame_1_errscaleX = obj.errscaleX + obj.Q_scaleX;
            frame_1_errscaleY = obj.errscaleY + obj.Q_scaleY;
            frame_1_errthetha = obj.errthetha + obj.Q_thetha;
            frame_1_errtransX = obj.errtransX + obj.Q_transX;
            frame_1_errtransY = obj.errtransY + obj.Q_transY;

            % 计算卡尔曼增益: K = P / (P + R), where P is predicted covariance, R is measurement noise
            gain_scaleX = frame_1_errscaleX / (frame_1_errscaleX + obj.R_scaleX);
            gain_scaleY = frame_1_errscaleY / (frame_1_errscaleY + obj.R_scaleY);
            gain_thetha = frame_1_errthetha / (frame_1_errthetha + obj.R_thetha);
            gain_transX = frame_1_errtransX / (frame_1_errtransX + obj.R_transX);
            gain_transY = frame_1_errtransY / (frame_1_errtransY + obj.R_transY);

            % 更新状态估计: new_state = predicted + K * (measurement - predicted)
            % Measurements are the raw accumulated sums (obj.sum_xxx).
            obj.scaleX = frame_1_scaleX + gain_scaleX * (obj.sum_scaleX - frame_1_scaleX);
            obj.scaleY = frame_1_scaleY + gain_scaleY * (obj.sum_scaleY - frame_1_scaleY);
            obj.thetha = frame_1_thetha + gain_thetha * (obj.sum_thetha - frame_1_thetha);
            obj.transX = frame_1_transX + gain_transX * (obj.sum_transX - frame_1_transX);
            obj.transY = frame_1_transY + gain_transY * (obj.sum_transY - frame_1_transY);

            % 更新误差协方差矩阵 new_P = (1 - K) * predicted_P
            obj.errscaleX = (1 - gain_scaleX) * frame_1_errscaleX;
            obj.errscaleY = (1 - gain_scaleY) * frame_1_errscaleY;
            obj.errthetha = (1 - gain_thetha) * frame_1_errthetha;
            obj.errtransX = (1 - gain_transX) * frame_1_errtransX;
            obj.errtransY = (1 - gain_transY) * frame_1_errtransY;
        end
    end
end