Line data Source code
1 : /*
2 : * Copyright (c) 2016 The WebRTC project authors. All Rights Reserved.
3 : *
4 : * Use of this source code is governed by a BSD-style license
5 : * that can be found in the LICENSE file in the root of the source
6 : * tree. An additional intellectual property rights grant can be found
7 : * in the file PATENTS. All contributing project authors may
8 : * be found in the AUTHORS file in the root of the source tree.
9 : */
10 :
11 : #include <limits>
12 :
13 : #include "webrtc/base/checks.h"
14 : #include "webrtc/base/logging.h"
15 : #include "webrtc/base/timestampaligner.h"
16 : #include "webrtc/base/timeutils.h"
17 :
18 : namespace rtc {
19 :
20 0 : TimestampAligner::TimestampAligner()
21 : : frames_seen_(0),
22 : offset_us_(0),
23 : clip_bias_us_(0),
24 0 : prev_translated_time_us_(std::numeric_limits<int64_t>::min()) {}
25 :
26 0 : TimestampAligner::~TimestampAligner() {}
27 :
28 0 : int64_t TimestampAligner::TranslateTimestamp(int64_t camera_time_us,
29 : int64_t system_time_us) {
30 0 : return ClipTimestamp(
31 0 : camera_time_us + UpdateOffset(camera_time_us, system_time_us),
32 0 : system_time_us);
33 : }
34 :
35 0 : int64_t TimestampAligner::UpdateOffset(int64_t camera_time_us,
36 : int64_t system_time_us) {
37 : // Estimate the offset between system monotonic time and the capture
38 : // time from the camera. The camera is assumed to provide more
39 : // accurate timestamps than we get from the system time. But the
40 : // camera may use its own free-running clock with a large offset and
41 : // a small drift compared to the system clock. So the model is
42 : // basically
43 : //
44 : // y_k = c_0 + c_1 * x_k + v_k
45 : //
46 : // where x_k is the camera timestamp, believed to be accurate in its
47 : // own scale. y_k is our reading of the system clock. v_k is the
48 : // measurement noise, i.e., the delay from frame capture until the
49 : // system clock was read.
50 : //
51 : // It's possible to do (weighted) least-squares estimation of both
52 : // c_0 and c_1. Then we get the constants as c_1 = Cov(x,y) /
53 : // Var(x), and c_0 = mean(y) - c_1 * mean(x). Substituting this c_0,
54 : // we can rearrange the model as
55 : //
56 : // y_k = mean(y) + (x_k - mean(x)) + (c_1 - 1) * (x_k - mean(x)) + v_k
57 : //
58 : // Now if we use a weighted average which gradually forgets old
59 : // values, x_k - mean(x) is bounded, of the same order as the time
60 : // constant (and close to constant for a steady frame rate). In
61 : // addition, the frequency error |c_1 - 1| should be small. Cameras
62 : // with a frequency error up to 3000 ppm (3 ms drift per second)
63 : // have been observed, but frequency errors below 100 ppm could be
64 : // expected of any cheap crystal.
65 : //
66 : // Bottom line is that we ignore the c_1 term, and use only the estimator
67 : //
68 : // x_k + mean(y-x)
69 : //
70 : // where mean is plain averaging for initial samples, followed by
71 : // exponential averaging.
72 :
73 : // The input for averaging, y_k - x_k in the above notation.
74 0 : int64_t diff_us = system_time_us - camera_time_us;
75 : // The deviation from the current average.
76 0 : int64_t error_us = diff_us - offset_us_;
77 :
78 : // If the current difference is far from the currently estimated
79 : // offset, the filter is reset. This could happen, e.g., if the
80 : // camera clock is reset, or cameras are plugged in and out, or if
81 : // the application process is temporarily suspended. Expected to
82 : // happen for the very first timestamp (|frames_seen_| = 0). The
83 : // threshold of 300 ms should make this unlikely in normal
84 : // operation, and at the same time, converging gradually rather than
85 : // resetting the filter should be tolerable for jumps in camera time
86 : // below this threshold.
87 : static const int64_t kResetThresholdUs = 300000;
88 0 : if (std::abs(error_us) > kResetThresholdUs) {
89 0 : LOG(LS_INFO) << "Resetting timestamp translation after averaging "
90 0 : << frames_seen_ << " frames. Old offset: " << offset_us_
91 0 : << ", new offset: " << diff_us;
92 0 : frames_seen_ = 0;
93 0 : clip_bias_us_ = 0;
94 : }
95 :
96 : static const int kWindowSize = 100;
97 0 : if (frames_seen_ < kWindowSize) {
98 0 : ++frames_seen_;
99 : }
100 0 : offset_us_ += error_us / frames_seen_;
101 0 : return offset_us_;
102 : }
103 :
104 0 : int64_t TimestampAligner::ClipTimestamp(int64_t filtered_time_us,
105 : int64_t system_time_us) {
106 0 : const int64_t kMinFrameIntervalUs = rtc::kNumMicrosecsPerMillisec;
107 : // Clip to make sure we don't produce timestamps in the future.
108 0 : int64_t time_us = filtered_time_us - clip_bias_us_;
109 0 : if (time_us > system_time_us) {
110 0 : clip_bias_us_ += time_us - system_time_us;
111 0 : time_us = system_time_us;
112 : }
113 : // Make timestamps monotonic, with a minimum inter-frame interval of 1 ms.
114 0 : else if (time_us < prev_translated_time_us_ + kMinFrameIntervalUs) {
115 0 : time_us = prev_translated_time_us_ + kMinFrameIntervalUs;
116 0 : if (time_us > system_time_us) {
117 : // In the anomalous case that this function is called with values of
118 : // |system_time_us| less than |kMinFrameIntervalUs| apart, we may output
119 : // timestamps with with too short inter-frame interval. We may even return
120 : // duplicate timestamps in case this function is called several times with
121 : // exactly the same |system_time_us|.
122 0 : LOG(LS_WARNING) << "too short translated timestamp interval: "
123 0 : << "system time (us) = " << system_time_us
124 0 : << ", interval (us) = "
125 0 : << system_time_us - prev_translated_time_us_;
126 0 : time_us = system_time_us;
127 : }
128 : }
129 0 : RTC_DCHECK_GE(time_us, prev_translated_time_us_);
130 0 : RTC_DCHECK_LE(time_us, system_time_us);
131 0 : prev_translated_time_us_ = time_us;
132 0 : return time_us;
133 : }
134 :
135 : } // namespace rtc
|