nisse | 23e3049 | 2016-06-22 15:36:53 | [diff] [blame] | 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 "webrtc/base/logging.h" |
| 12 | #include "webrtc/base/timestampaligner.h" |
| 13 | |
| 14 | namespace rtc { |
| 15 | |
| 16 | TimestampAligner::TimestampAligner() : frames_seen_(0), offset_us_(0) {} |
| 17 | TimestampAligner::~TimestampAligner() {} |
| 18 | |
| 19 | int64_t TimestampAligner::UpdateOffset(int64_t camera_time_us, |
| 20 | int64_t system_time_us) { |
| 21 | // Estimate the offset between system monotonic time and the capture |
| 22 | // time from the camera. The camera is assumed to provide more |
| 23 | // accurate timestamps than we get from the system time. But the |
| 24 | // camera may use its own free-running clock with a large offset and |
| 25 | // a small drift compared to the system clock. So the model is |
| 26 | // basically |
| 27 | // |
| 28 | // y_k = c_0 + c_1 * x_k + v_k |
| 29 | // |
| 30 | // where x_k is the camera timestamp, believed to be accurate in its |
| 31 | // own scale. y_k is our reading of the system clock. v_k is the |
| 32 | // measurement noise, i.e., the delay from frame capture until the |
| 33 | // system clock was read. |
| 34 | // |
| 35 | // It's possible to do (weighted) least-squares estimation of both |
| 36 | // c_0 and c_1. Then we get the constants as c_1 = Cov(x,y) / |
| 37 | // Var(x), and c_0 = mean(y) - c_1 * mean(x). Substituting this c_0, |
| 38 | // we can rearrange the model as |
| 39 | // |
| 40 | // y_k = mean(y) + (x_k - mean(x)) + (c_1 - 1) * (x_k - mean(x)) + v_k |
| 41 | // |
| 42 | // Now if we use a weighted average which gradually forgets old |
| 43 | // values, x_k - mean(x) is bounded, of the same order as the time |
| 44 | // constant (and close to constant for a steady frame rate). In |
| 45 | // addition, the frequency error |c_1 - 1| should be small. Cameras |
| 46 | // with a frequency error up to 3000 ppm (3 ms drift per second) |
| 47 | // have been observed, but frequency errors below 100 ppm could be |
| 48 | // expected of any cheap crystal. |
| 49 | // |
| 50 | // Bottom line is that we ignore the c_1 term, and use only the estimator |
| 51 | // |
| 52 | // x_k + mean(y-x) |
| 53 | // |
| 54 | // where mean is plain averaging for initial samples, followed by |
| 55 | // exponential averaging. |
| 56 | |
| 57 | // The input for averaging, y_k - x_k in the above notation. |
| 58 | int64_t diff_us = system_time_us - camera_time_us; |
| 59 | // The deviation from the current average. |
| 60 | int64_t error_us = diff_us - offset_us_; |
| 61 | |
| 62 | // If the current difference is far from the currently estimated |
| 63 | // offset, the filter is reset. This could happen, e.g., if the |
| 64 | // camera clock is reset, or cameras are plugged in and out, or if |
| 65 | // the application process is temporarily suspended. The limit of |
| 66 | // 300 ms should make this unlikely in normal operation, and at the |
| 67 | // same time, converging gradually rather than resetting the filter |
| 68 | // should be tolerable for jumps in camera time below this |
| 69 | // threshold. |
| 70 | static const int64_t kResetLimitUs = 300000; |
| 71 | if (std::abs(error_us) > kResetLimitUs) { |
| 72 | LOG(LS_INFO) << "Resetting timestamp translation after averaging " |
| 73 | << frames_seen_ << " frames. Old offset: " << offset_us_ |
| 74 | << ", new offset: " << diff_us; |
| 75 | frames_seen_ = 0; |
| 76 | prev_translated_time_us_ = rtc::Optional<int64_t>(); |
| 77 | } |
| 78 | |
| 79 | static const int kWindowSize = 100; |
| 80 | if (frames_seen_ < kWindowSize) { |
| 81 | ++frames_seen_; |
| 82 | } |
| 83 | offset_us_ += error_us / frames_seen_; |
| 84 | return offset_us_; |
| 85 | } |
| 86 | |
| 87 | int64_t TimestampAligner::ClipTimestamp(int64_t time_us, |
| 88 | int64_t system_time_us) { |
| 89 | // Make timestamps monotonic. |
| 90 | if (!prev_translated_time_us_) { |
| 91 | // Initialize. |
| 92 | clip_bias_us_ = 0; |
| 93 | } else if (time_us < *prev_translated_time_us_) { |
| 94 | time_us = *prev_translated_time_us_; |
| 95 | } |
| 96 | |
| 97 | // Clip to make sure we don't produce time stamps in the future. |
| 98 | time_us -= clip_bias_us_; |
| 99 | if (time_us > system_time_us) { |
| 100 | clip_bias_us_ += time_us - system_time_us; |
| 101 | time_us = system_time_us; |
| 102 | } |
| 103 | prev_translated_time_us_ = rtc::Optional<int64_t>(time_us); |
| 104 | return time_us; |
| 105 | } |
| 106 | |
| 107 | } // namespace rtc |