J Forensic Sci, 2019

doi: 10.1111/1556-4029.14053
CASE REPORT Available online at: onlinelibrary.wiley.com

DIGITAL & MULTIMEDIA SCIENCES

Brandon Epstein ,1 B.S.; and Bryce Garreth Westlake ,2 Ph.D.

Determination of Vehicle Speed from

Recorded Video Using Reverse Projection
Photogrammetry and File Metadata†

ABSTRACT: The prevalence of security and in-car video has increased the number of motor vehicle accidents captured on digital video.
However, inconsistencies in how to accurately determine time and distance for vehicle speed has led to examinations with varying results. A
potential solution for calculating time intervals is to use frame timing contained within many digital video file’s metadata, recorded with
0.000001 sec precision. This paper examines a fatal motor vehicle accident where frame timing information was used with distance measure-
ments from reverse projection photogrammetry to calculate vehicle speed. A margin of error was then calculated based on the accuracy in per-
forming reverse projection photogrammetry distance measurements. The resulting speed calculation was then compared to event data recorder
data and found to be within an average of
1.43538 MPH. Using specific time intervals may lead investigators to more accurate speed calcula-
tions, specifically those involving variable frame rate video.

KEYWORDS: forensic science, speed calculation, forensic video analysis, frame timing, reverse projection photogrammetry, accident recon-
struction, video metadata

The prevalence of closed-circuit television (CCTV), or video the time no accident reconstruction was conducted, however one
surveillance, in crime detection and prevention has necessitated of the vehicles, a blue Dodge Ram, was equipped with an event
investigators to be able to accurately analyze attributes present data recorder (EDR)—essentially a vehicle’s black box—that
within digital recordings. A commonly conducted analysis is to recorded the vehicle’s speed, and other parameters, prior to
determine the speed of a recorded object. In law enforcement impact. The restaurant’s exterior security video camera also cap-
use of force cases, this may mean the rate/speed an object/person tured the collision; that recorded video was used to calculate the
is moving toward/away from an officer (1). It can also be used Dodge Ram’s speed prior to impact. The Middlesex County
to determine the amount of time that has elapsed between New Jersey Prosecutor’s Office (MCPO) requested that the
events, such as how long after a person reached for an object speed of the blue Dodge pickup truck, prior to impact, be deter-
was a gun fired. In motor vehicle accident cases, analysis of mined from the video footage of the accident. It should be noted
video may be needed to determine a vehicle’s speed, when an that although the examiner was informed that investigators pos-
accident reconstruction was unable to be conducted or there was sessed speed information from the EDR, those speeds were not
insufficient evidence at the scene (e.g., skid marks) to draw a disclosed until after speed was determined from the security
conclusion (2). In determining speed within a digital video video.
recording (DVR), it is imperative that methods for accurately The examiner was provided with a.MP4 video file from the
calculating both the time interval and distance travelled be accu- Wonder Seafood Restaurant DVR by the MCPO investigator.
rately identified by investigators. Upon initial review of the file, it was found to have a resolution
of 1280x720 pixels encoded in a H.264 format. Initial metadata
Case Report analysis in FFmpeg (v 3.0.6) (3) and MediaInfo (v 0.7.90)
reported the video to have a frame rate of 25 frames per second
On July 25, 2016 a fatal accident occurred in front of the (FPS). The examiner believed this to be incorrect, as 25 FPS is
Wonder Seafood Restaurant at 1984 RT-27 in Edison, NJ. At known to be reported in various media interrogation tools when
frame rate cannot be determined. Additionally, the DVR sys-
1 tem’s settings and owner’s manual both indicated that the DVR
Middlesex County Prosecutor’s Office, 25 Kirkpatrick Street, 3rd Floor,
New Brunswick, NJ 08901. was recording at 30 FPS. A visual inspection of the video file
2
Department of Justice Studies, San Jose State University, San Jose, CA found that the incident occurred almost directly in front of the
95192. camera, with the vehicle in question perpendicular to the camera
Corresponding author: Brandon Epstein, B.S. E-mail: brandon.epstein@co. travelling from left to right in the horizontal axis. Furthermore,
middlesex.nj.us
†
Presented at the 71st Annual Scientific Meeting of the American Acad-
the incident occurred during the day, in good lighting conditions,
emy of Forensic Sciences, February 18-23, 2019, in Baltimore, MD. and without any obstruction of the vehicles.
Received 20 Dec. 2018; and in revised form 5 Mar. 2019; accepted 6 An initial analysis of the video file was conducted to identify
Mar. 2019. video frames to be utilized in the examination. This was

© 2019 American Academy of Forensic Sciences 1

2 JOURNAL OF FORENSIC SCIENCES

accomplished using a visual macroblock analysis of the file gen-

Results
erated by FFmpeg (3). The analysis was conducted to identify
frames of the video prior to impact that consisted of newly Reverse Projection Distance Measurements
encoded information in order to accurately place the Dodge Ram
Reverse projection photogrammetry involves the positioning
at a position on the roadway. A total of seven frames were iden-
of a camera and recording in the perspective and aspect ratio
tified for use in the examination (Fig. 1). The frame immediately
duplicating the original imagery. A calibrated measuring device
prior to impact was not utilized as it consisted of too much pre-
may then be used to complete the requested analysis (5). Rev-
dictive information for the examiner to accurately determined the
erse projection photogrammetry has been shown to effectively
location of each vehicle (4).
measure objects and distances in images (6–8). In this case,
A combination of reverse projection photogrammetry and
investigators returned to the scene of the incident to employ
frame timing analysis was utilized to calculate speed as distance
reverse projection, to determine the distance the Dodge Ram
over time. The reverse projection and frame timing analysis were
travelled prior to impact.
viewed as two separate parts of the examination whose results
Upon returning to the scene, investigators found a LTS model
would be used to calculate speed. That calculated speed would
LTD8308T-FT digital video recorder connected to eight analog
be verified against the vehicle’s EDR data.
cameras. Building management stated that the DVR settings and

FIG. 1––Seven frames selected from macroblock analysis.

 EPSTEIN AND WESTLAKE . SPEED CALCULATION IN RECORDED DIGITAL VIDEO 3

all associated cameras have not been changed or accessed since found that the difference in frame display time varied from
the initial incident. Furthermore, the system settings displayed 0.03136–0.06666 sec as opposed to a constant 0.04000 sec for
by the DVR matched the same resolution and codec found on 25 FPS video or 0.03333 sec for 30 FPS video. The differences
the initially acquired video file. between pkt_pts_time for each frame used in the reverse projec-
The DVR system’s HDMI out signal was then connected to tion were noted and then used to calculate the time it took for
an Epiphan DVI2USB video grabber in order to capture live the vehicle to travel between identified frames (Table 3). For
video from the system. This allowed the investigator to view the example, in Table 2, the pkt_pts_time for Point 1 was
video stream as it was displayed, from the same camera that 86995.12060 and 86995.58731 for Point 4. Therefore, the time
recorded the initial incident, using the same system settings. The it took for the vehicle to travel from Point 1 to Point 4, noted in
live video displayed from the restaurant’s DVR system was then Table 3, was the difference between each’s pkt_pts_time:
opened within Amped FIVE forensic video software (revision 0.46671 sec.
8678). Using the software’s video mixer function that overlays
the live video feed from the Epiphan DVI2USB video grabber
Speed Determination
over the recorded footage, it was also confirmed by visual
inspection that static elements of the images (i.e., parking and Having identified the distances the vehicle travelled (in feet)
lane line striping, curbs, and exterior signage) remained in the in Table 1, and the time that it took to travel those distances
same locations; indicating that the camera had not moved since (in seconds) in Table 2, speed was calculated as distance over
the time of the incident. The live feed from the Epiphan time. That feet per second measurement was then converted
DVI2USB video grabber was then overlaid on still images of to miles per hour (Table 4). The final calculated speed
the seven identified pertinent frames. represents the average speed travelled between two points, not
Using the overlaid recorded images and the live feed from the at any exact point in time. Using the calculated speed
Epiphan DVI2USB video grabber as a guide, an investigator between multiple points, acceleration and deceleration can be
with a reference marker was given instructions by two-way radio evaluated (Fig. 3).
to mark the roadway in locations that corresponded to the same
point on the vehicle’s front bumper in each identified frame
Margin of Error
(Fig. 2) (5). Seven marks identified as points 1–5 and impact
points 1 and 2 were used to determine the distance that the vehi- Due to the spatial (interframe) compression within the frames
cle travelled. Distances were then calculated between points 1 of the recorded video, as well as a potential for human error dur-
and 5 and impact points 1 and 2 as well as points 1–3 and ing the reverse projection distance determination, a margin of
points 4 and 5 (Table 1). error for the calculated speed was determined. This was accom-
plished by calculating the physical dimensions of a pixel found
approximately the same distance from the camera as the Dodge
Frame Timing
Ram. In this case, a roadway lane stripe was measured on scene
The MP4 video file from the Wonder Seafood Restaurant and found to be 100 1″ long; this same stripe was then measured
DVR was analyzed to determine specific frame timing differ- in Amped FIVE to be 115 pixels. The resultant calculation
ences. FFprobe was utilized to create a frame analysis spread- found the physical length of each pixel to be 13/64 inches. Given
sheet consisting of the individual presentation times for each that the selected stripe was slightly further from the camera than
video frame (3). While the frame analysis spreadsheet consists the Dodge Ram, the pixel measurement is actually larger than
of many different metadata values for video files, the packet pre- where the vehicle travelled, resulting in a more conservative
sentation time (pkt_pts_time) was utilized, as it is derived from margin of error.
the encoding time of a frame to the 0.000001 sec. Considering the inability to place a specific real-world point
The FFprobe frame analysis derives the decode and presenta- within a compressed image, a six-pixel (65/16 -inch) margin of
tion times from the video file’s container, whereas the initial error was utilized in the reverse projection distance calculations.
metadata analysis looked at the video stream or codec for this This margin of error was used to calculate a maximum and mini-
information. As decode and presentation time in the file con- mum speed for calculated points (Figs 3, 4 and 5). This led to
tainer is an integral part of the ITU-T H.264 standard (9) it may an error margin of
1.10672 MPH (p < 0.05). However, it is
be more reliable than any attempts to decode it within the video worth noting that because speed calculation is an average of dis-
stream. FFprobe reports packet decode time as “pkt_dts_time” tance over time, the greater the distance measured between two
and packet presentation time as “pkt_pts_time”. Packet decode points, the smaller the margin of error becomes.
time is the time at which the frame is intended to be decoded,
whereas packet presentation time is the time at which it is meant
Comparison with EDR Data
to be displayed. In video files that contain bidirectional frames,
p-frames that are displayed after b-frames are actually decoded After completion of all calculations, the examiner received the
before those b-frames as the information needed to display the EDR data from the vehicle (Table 5). The calculated speed from
b-frame is contained in the p-frame after it. In this case investi- the Dodge Ram was then compared to the acquired Bosch EDR
gation, the examiner found a variable frame rate video, yet was data. The Bosch EDR records a speed at 0.1 sec intervals prior
able to determine elapsed time using the time each frame was to impact; it does not record distance. Using the visual display
intended to be decoded and presented using the FFprobe report of the vehicles’ point of impact, time was counted back to corre-
(Table 2). The packet presentation and packet decode times were late the EDR data with the calculated speed. Additionally, the
also identical as there were no bidirectional frames present in the EDR data reports speed at a specific point in time, whereas the
file. calculated speed is an average over a distance. For this reason,
Although the initial metadata analysis using FFmpeg and the EDR speeds for the matching distances were determined and
MediaInfo found the file to be 25 FPS, the FFprobe spreadsheet compared to the calculated speed as well. Put another way,
4 JOURNAL OF FORENSIC SCIENCES

FIG. 2––Reverse projection roadway marking.

because the EDR speed is recorded every 0.1 sec, the closest a more complex endeavor because of challenges in properly
two visual displayed values were averaged to show average identifying a video’s frame rate.
speed over time. The comparison of EDR average speed and cal- When using video playback or editing applications, a common
culated vehicle speed to impact point 2 found an average differ- method for calculating time differences is to overlay the
ence of 1.43538 MPH (Table 5), with a margin of error of recorded video’s timecode on the footage, to display the time of

1.17114 MPH (p < 0.05). individual frames. This timecode can be derived from the video
file itself or derived from settings within the software used for
video playback. Due to the large number of software applica-
Discussion
tions available to view and edit digital video, this means that
Traffic accident reconstructionists and video practitioners are there are countless ways in which timecodes can be derived. As
often tasked with calculating a vehicle’s speed from a recorded a result, there are inherent issues when attempting to determine
video. Generally, this involves determining the time elapsed for frame timing using the video timecode. Often, security digital
a vehicle to travel between two points. Various methods have video recorders (DVRs), and other camera systems, record at
been used for determining distance travelled, with varying levels frame rates other than the standard 29.97 frames per second
of precision and accuracy (10,11). However, calculating time is found in NTSC television. Many video playback and editing
 EPSTEIN AND WESTLAKE . SPEED CALCULATION IN RECORDED DIGITAL VIDEO 5

FIG. 3––Average calculated speed from point 1 with margin of error (in MPH).

TABLE 1––Reverse projection distance measurements. TABLE 4––Calculated average speed.

Point 4 Point 5 Impact Point 1 Impact Point 2 Point 4 Point 5 Impact 1 Impact 2
Point 1 440 1/8″ 460 8½″ 730 1½″ 750 8″ Point 1 64.29495 63.68807 62.31931 61.90412
Point 2 310 1/8″ 330 7″ 600 1″ 620 8″ Point 2 63.42722 62.44605 61.44761 59.08752
Point 3 210 6″ 240 2″ 500 7½″ 530 3½″ Point 3 62.82017 61.78668 60.91098 60.55466
Point 4 290 1″ 290 8″ Point 4 59.48921 55.16177
Point 5 260 6″ 290 0″ Point 5 60.22726 59.31342

A common method for determining frame rate is to evaluate

TABLE 2––Pertinent frames with packet presentation time (pkt_pts_time) the video file’s metadata. There are open-source software tools
and packet decode time (pkt_dts_time). that will read the metadata of proprietary video files and report
several file attributes, including frame rate. While the reported
Frame pkt_pts_time pkt_dts_time frame rate is sufficient to enable playback, it may not account
Point 1 86995.12060 86995.12060 for the exact timing between displayed frames, particularly with
Point 2 86995.25396 86995.25396 variable frame rate video files. This inaccuracy can be important
Point 3 86995.35396 86995.35396 when, for example, attempting to determine speed of a vehicle
Point 4 86995.58731 86995.58731 in a fatal accident.
Point 5 86995.62064 86995.62064
Impact 1 86995.92064 86995.92064
Neither playback applications nor metadata frame rate evalua-
Impact 2 86995.95400 86995.95400 tion accounts for the specific elapsed time between video frames.
Rather they identify the number of frames displayed over time,
reported as frames per second. While a video file may have a
frame rate of 30 FPS, each individual frame may not be dis-
played at equal 0.03333 sec intervals. Given the nature of
TABLE 3––Pertinent frame time differences.
DVRs, many may not be able to encode or store every frame of
Point 4 Point 5 Impact 1 Impact 2 video, resulting in dropped frames or repeated (padded) frames.
Identification of missing or repeated video frames has been
Point 1 0.46671 0.50004 0.80004 0.83340 addressed in previous research using visual cues to calculate
Point 2 0.33335 0.36668 0.66668 0.70004
Point 3 0.23335 0.26668 0.56668 0.60004 frame timing in recorded video (14,15). However, this method
Point 4 0.33333 0.36668 may not account for specific frame differences in variable frame
Point 5 0.30000 0.33335 rate video, common within CCTV video, resulting in a regular
or irregular pattern of frame time differences throughout a video
file. It may also lack precision due to visual limitations when
software applications are not equipped to properly understand examining those cues, like smaller resolutions, motion blur, and
this timecode, and simply default to 29.97 FPS timecode (or high levels of compression, common within security DVRs.
another frame rate) (12,13). Therefore, the resultant timecode In order to determine the precise difference in time between
displayed could lead to a misinterpretation of the time elapsed displayed frames in video files, an analysis of each frame’s pre-
between frames. sentation time can be conducted using file metadata. Denoted
6 JOURNAL OF FORENSIC SCIENCES

FIG. 4––Average calculated speed from point 2 with margin of error (in MPH).

FIG. 5––Average calculated speed from point 3 with margin of error (in MPH).

specific timestamps for the time each frame is to be presented

TABLE 5––Calculated average speed to impact point 2 with EDR speed (in
MPH). (displayed) is recorded in the file’s metadata; the ability to
decode this presentation timestamp is dependent on the file con-
Calculated EDR Average Delta tainer. This packet presentation time will denote the specific time
for when each frame is intended to be displayed, often to the
Point 1-Impact 2 61.90412 60.25 1.65412
Point 2-Impact 2 59.08752 59.571 0.48348 0.00001 sec. The analysis of packet presentation time can be
Point 3-Impact 2 60.55466 59 1.55466 more accurate and precise than the existing methods as it is
Point 4-Impact 2 55.16177 58 2.83823 specific to each individual frame. The presentation time can also
Point 5-Impact 2 59.31342 58.667 0.64642 account for variable frame rates, dropped or padded frames, as
well as small timing differences in the encoding process.

within the video file metadata as “packet presentation time,” the

Conclusion
presentation time of each frame is an integral part of the ITU-T
H.264 encoding standard, as such it will be present in CCTV Understanding frame timing within video files is critical to
video with H.264 encoding (8). By the nature of the standard, accurately determining vehicle speed. Using FFmpeg to identify
 EPSTEIN AND WESTLAKE . SPEED CALCULATION IN RECORDED DIGITAL VIDEO 7

TABLE 6––Average calculated speed to impact point 2 with frame rates 25 small task to perform a reverse projection and understand the
FPS and 30 FPS (in MPH). intricacies in frame timing differences. These projects are often
attempted by traffic accident reconstructionists with little to no
Calculated 25 FPS 30 FPS digital video training or experience. Future studies regarding the
Point 1-Impact 2 61.90412 61.41774 73.70865 ability for personnel to conduct this examination with varying
Point 2-Impact 2 59.08752 60.82886 73.00194 levels of training would provide insight as to the discipline or
Point 3-Impact 2 60.55466 60.55870 72.67771 amount of training required to accurately complete the task.
Point 4-Impact 2 55.16177 56.18686 67.43098
Point 5-Impact 2 59.31342 61.78976 74.15512
Acknowledgments

specific frame intervals from the recorded video allowed the The authors would like to acknowledge Bertram Lyons and
examiner to calculate speed from variable frame rate video more Blake Sawyer, as well as the rest of the Scientific Working
accurately than using an average of frames per second. Had the Group on Digital Evidence Video Committee, for their contribu-
examination been conducted using an incorrect frame rate of 25 tions to a deeper understanding of packet presentation time in
or 30 FPS, the resulting speed calculation could have been as digital video files.
much as 13.91442 MPH incorrect (Table 6). This variance fur-
ther demonstrates the need to effectively determine frame timing References
using packet presentation time in these examinations. The avail-
1. State of Texas v. Derick Wiley, 2018.
ability of EDR data in this investigation was integral in validat- 2. State of New Hampshire v. Witty, 2018.
ing the calculated speed, demonstrating that the frame timing 3. SWGDE. SWGDE technical notes on FFmpeg. https://www.swgde.org/d
and speed measurements in this examination were accurate. ocuments/Current%20Documents/SWGDE%20Technical%20Notes%20on
An additional method to validate these findings can be con- %20FFmpeg (accessed November 20, 2018).
ducted by driving a vehicle at a known speed through the same 4. SWGDE. SWGDE technical overview of digital video files. https://www.
swgde.org/documents/Current%20Documents/SWGDE%20Technical%
scene and calculating that speed from the recorded video. This 20Overview%20of%20Digital%20Video%20Files (accessed November
was not completed in this examination as the lowest calculated 20, 2018).
speed was still 11 MPH greater than the speed limit of the road- 5. SWGDE. SWGDE best practices for the forensic use of photogrammetry.
way. Because the roadway in question is highly travelled, and a https://www.swgde.org/documents/Current%20Documents/SWGDE%
20Best%20Practices%20for%20the%20Forensic%20Use%20of%20Pho
major route to two hospitals, a decision was made to minimize togrammetry (accessed November 20, 2018).
any potential traffic delays by closing the road to perform this 6. Hoogeboom B, Alberink I, Vrijdak D. Photogrammetry in digital foren-
task. It is also a challenge to accurately record the known speed sics. In: Ho TS, Shujun L, editors. Handbook of digital forensics of mul-
of the vehicle at the time it is recorded. Even with a calibrated timedia data and devices. Chichester: Wiley, 2015;183–218.
speedometer, most vehicles have an analog speed display which 7. Hoogeboom B, Alberink I. Measurement uncertainty when estimating the
velocity of an allegedly speeding vehicle from images. J Forensic Sci
makes determining exact speed troublesome. Even when 2010;55(5):1347–51.
recorded, it is only precise to one MPH. The use of a radar/laser 8. Gebruik van 3D-modellen in forensisch onderzoek [Use of 3D models in
speed measurement device or other GPS/GLONASS device to forensic investigation]. Den Haag, Nederland: Ministerie van Veiligheid
measure speed of the known vehicle may help increase accuracy. en Justite, 2014.
9. International Telecommunications Union. Series H: audiovisual and
There is need for additional research in determining the multimedia systems: infrastructure of audiovisual services–coding of
elapsed time between video frames using reported packet presen- moving video. Geneva, CH: International Telecommunications Union,
tation times in various digital media files. This applies not only 2017.
to H.264 compressed files in general, but also to different imple- 10. Lue S, Yang X, Cui J, Yin Z. A novel pixel-based method to estimate
mentations of the standard. A wider study of different manufac- the instantaneous velocity of a vehicle from CCTV images. J Forensic
Sci 2017;62(4):1071–4.
turer’s DVRs, using different frame rate setting and variable 11. Han I. Car speed estimation based on cross-ratio using video data of car-
frame rates, could provide additional data to gain a better under- mounted camera (black box). Forensic Sci Int 2016;269:89–96.
standing of how reliable packet presentation time is. 12. Adobe. Adobe Premiere Pro CC help. https://helpx.adobe.com/pdf/premie
Research on reverse projection photogrammetry and the accu- re_pro_reference.pdf (accessed December 15, 2018)
13. Apple. Final Cut Pro X user guide. https://manuals.info.apple.com/MAN
racy, as compared to LiDAR crime scene scanning, could also UALS/1000/MA1666/en_US/final_cut_pro_x-10.1.2-user_guide.pdf (ac-
help to improve accuracy (16). This could help to reduce the cessed December 15, 2018).
margin of error in these calculations. Additional black box stud- 14. Cheng YK, Wong KH, Tao CH, Tam YY, Tsang CN, Poon KC. Cali-
ies into an examiner’s ability to accurately use LiDAR crime bration of dashboard cameras for speed determination from video record-
scene scanning, reverse projection, and/or other methods of pho- ing. Impact 2016;24(3):18–25.
15. Lower S, Stevens R, Crouch M, Cash S. Collision investigation:
togrammetry could also assist in determining the accuracy of CCTV playback and validation using a lightboard. Impact 2017;
vehicle speed calculations. 25(3):20–7.
Additional considerations should be given to the ability of 16. Meline K, Bruehs W. A comparison of reverse projection and laser scan-
examiners to complete an examination of this nature. It is no ning photogrammetry. J Forensic Identif 2018;68(2):281–92.