nvtk_mp42gpx.py revisited – now with TS support

It appears that not only my quick hack of a script is popular but also that the dashcam manufacturers keep changing things.

In this case the biggest change (beyond utterly stupid data obfuscation) is the switch to the TS format.

In retrospect TS format is an obvious choice for something like dashcam as it is very resilient to crashes (no pun intended).

TS format unlike the MP4/MOV format is very simple, it is pretty much made up of fixed 188 byte segments which have ASCII ‘G’ as a header. The actual payload lives in last 184 bytes while the first 4 bytes are reserved for header (of course this is all simplified).

The biggest difficulty of adding TS functionality is that Blueskysea B4K camera was engineered by sadists and they split the payload into two packets (arbitrary cutting off at the seconds 4 bytes). If you are the person at the BlueSkysea who made this obfuscation – fuck you.

In anyway this update is massive rewrite.
Just to make it easy here is the script: nvtk_mp42gpx.py

2020 update of the nvtk_mp42gpx.py script

Surprisingly to me the script that I haphazardly put together turned out to be very popular.
I decided to quickly update it (after multiple feature and bug fixes requests)…

The highlights are:

  • Automatic finding of the correct position of the data block (more on that below).
  • Introduction of -m flag, which creates separete GPX file per input file (default behaviour is a single GPX output file for multiple input files).
  • More robust argument parsing (now it should always display help if wrong arguments are used)
  • Introduction of -d flag, which allows to de-obfuscate GPS coordinates.
Continue reading 2020 update of the nvtk_mp42gpx.py script

ILDVR – and their lack of professionalism

It all started with me purchasing ILDVR INC-MH40D06 IP Camera. I decided to poke at it and discovered some interesting and blatant security flaws.

About a year ago I contacted ILDVR (Arnold and Marika Wei) regarding the security issues, which got no response.
After about a year of the camera sitting on a shelf, I decided to poke at it again.
Which prompted me to send them this email:

For which I got a friendly response from Marika:

To which I replied, asking for firmware update (which I thought was reasonable to expect firmware updates for products with serious security flaws):

The only response I got is this peculiar email from sales@ildvr.com:

So, it seems that:
1) ILDVR.com/ILDVR does not care about security
2) ILDVR.com/ILDVR does not care about PR
3) ILDVR.com/ILDVR does not care about customers

Perhaps they should adopt the following motto:

“GO TO HELL! – ILDVR (where security does not matter)”.

To be honest, I would probably let go this whole thing if they simply not responded. It would have taken them less effort to not to respond either. Instead they chose to send me email with “GO TO HELL!”. I find this thing very hilarious.

It is even more hilarious if you look at google search results:

ILDVR INC-MH40D06 security nightmare part 2

I have put off the ILDVR camera, as I kind of lost interest.
For previous posts see here, here and here.

I was bored so I decided to poke at again.

I was interested where does the camera store users and in what format. What I found out is an atrocious mocking of security.
The camera stores local users and their passwords (in plain text) in following file:


Yep: the same directory which is accessible without auth via port 10081. So if you forgot password (and forgot the silly hardcoded HANKVISION), then you can get a reminder what it is by simply going here:


There is also another “binary” file that contains interesting references to HANKVISION and local users:


strings that and you get following:


Changing the password to HANKVISION reveals that this string is “encoded” “HANKVISION”:


Not sure if the obfuscation is worth spending time on, especially when we already know HANKVISION is hardcoded in web server binary and OwnUserInfo.txt already contains passwords in clear texts.

strace-ing ‘webs’ process during certain conditions opens the /tmp/umconfig.txt, which contains following:


The “hashes” correlate to /mnt/flash/data/UserInfo…

Looking firmware upload function (in browse/javascript/sysInf.js) I found this bit:

function fileUpload(){
		var typeAllow = [".ifu", "macaddr.txt", "deviceid.txt", "sn.txt", "audio.dat", ".bin", ".png", ".ifc", ".lib", ".uid", ".pid","logo.gif","whitelist.txt"];
		var fileType = ["ifu", "mac", "deviceid", "sn", "audio", "bin", "png", "ifc", "lib", "uid", "pid","gif","wlst"];

I have tested the upload function with logo.gif and that worked: the logo on top got replaced, so it brings a possibility of doing something more (sneaking in a binary?).

Looking at ‘webs’ binary I decided to google for strings in case someone leaked the source or these bastards stole somebody else’s work.
Here what I found:
The string:

webs: websWrite lost data, buffer overflow

Matches suspiciously named file here:

Same could be said for these strings:

webs: Listening for HTTP requests at address %s
webs: accept request

What is surprising is that they avoided doing execve calls where they could. IP addresses, routes, all set via ioctl, even time is set via settimeofday function. This removed possibility of command injection.

Here is what I believe is going on with this firmware:

The video side and core functionality has been lifted off SDK by Hisilicon. The web server stuff has been implemented by actual Hankvision people, most likely low paid undergraduate Chinese students. The core web server functionality has been lifted off the internet (see above).

What could have been done better without spending much on development:

Remove hard coded passwords!
Throw away all activeX crap (use MJPEG stream for “preview”).
Turn off telnet and leave ssh on with configurable password (perhaps make it a separate user?).
Do not store plain passwrods anywhere
Throw away all the dyndns and cloud nonsense.
Add actual off checkbox for FTP, Mail and SIP stuff (and possbly throw away SIP stuff).
Add VLC plug-in functionality.
Remove web server that listens on port 10081 exposing whole bunch of private data.

I am not sure what they are trying to achieve by not allowing SSH/Telnet access, but this is counter productive. I will not buy a security product to which I do not have control! Besides if I wanted to get access to your firmware, I don’t need SSH or Telnet, when I have RS232 and soldering iron.

For those who purchased this camera, if you really have to use it do the following:

Hexedit webs binary and change the HANKVISION bit to something else


Remove gateway setting (set it the same IP as camera) and preferably isolate camera from rest of the network (separate VLAN and port forwarding to recorder).


Just chuck it in the bin and never purchase anything from ILDVR again.

Shame on you ILDVR for not responding to me when I contacted you almost a year ago about hard coded passwords. Shame on you ILDVR for not providing root password or firmware updates.

python OpenCV basic motion detection

Here I will describe how I use OpenCV for capturing RTSP streams, with purpose of motion detection.

For basic OpenCV I use these two libraries:

import cv2
import numpy as np

cv2 is OpenCV library (second version), and numpy is python numeric lybrary (useful for manipulating matrices among other things).

To initiate capture one simply does following:

cap = cv2.VideoCapture('rtsp://')

In this example I use second stream (of lower resolution) for motion detection.

From there you can get heigh and width of the frame (this will be useful later):

width = cap.get(3)
height = cap.get(4)

I use BackgroundSubtractorMOG for motion detection (somewhat cheating ;)):

bg = cv2.BackgroundSubtractorMOG(100,3,0.6,30)

The magic is in parameters, I used following:
100 – history
3 – number of Gaussian mixtures
0.6 – background ratio
30 – noise strength
The numbers above are not necessarily “correct” but I came to them with error and trial (and “guestimation”).
Here is document in detail describing this algorithm: http://personal.ee.surrey.ac.uk/Personal/R.Bowden/publications/avbs01/avbs01.pdf

The actual capture loop looks something like this:

    ret, frame = cap.read()
    motion = bg.apply(frame, learningRate=0.005)
    kernel = np.ones((3, 3), np.uint8)
    motion = cv2.morphologyEx(motion, cv2.MORPH_CLOSE, kernel, iterations=1)
    motion = cv2.morphologyEx(motion, cv2.MORPH_OPEN, kernel, iterations=1)
    motion = cv2.dilate(motion,kernel,iterations = 1)
    contours, hierarchy = cv2.findContours(motion, cv2.RETR_EXTERNAL,cv2.CHAIN_APPROX_SIMPLE)
ret, frame = cap.read()

captures a single frame

motion = bg.apply(frame, learningRate=0.005)

extracts a black and white image with the background removed (learnignRate value has been chosed by error and trial).

Next four lines simply manipulate extracted image in such that it does following:
MORPH_CLOSE: removes small holes (up to 3×3 pixel, defined by kernel) within the object (“white”) in the extracted motion matrix.
MORPH_OPEN: removes small dots within the “background” (black) in the extracted motion matrix.
dilate: is making sure there all adjacent islands are joined together, so when we extract contours we get small amount of contours as result.

The “3×3 pixel” block comes from here:

kernel = np.ones((3, 3), np.uint8

The last step from processing frame is extracting the contours:

contours, hierarchy = cv2.findContours(motion, cv2.RETR_EXTERNAL,cv2.CHAIN_APPROX_SIMPLE)

The extracted contours can then be iterated and hull drawn around them:

for cnt in contours:
    hull = cv2.convexHull(cnt)

The hull points then can be checked against the mask if motion is inside of the area of interest:

Lets define the mask as the whole frame (I am pretty sure there is a better way;)):

mask_points = [
( 0 , 0 ),
( 1 , 0 ),
( 1 , 1 ),
( 0 , 1 ),

for point in mask_points:
    mask_array.append([[int(point[0] * width ), int(point[1] * height )]])
mask = np.array(mask_array, np.int32)

This looks cumbersome, but what I am achieving here is converting mask_points list of human readable relative coordinate tuples (eg: centre will be at (0.5,0.5)). Mask can be defined as a polygon with relative positioning of each corner to the frame (independent from pixel size).

We check if hull point is inside our mask

for point in hull:
    distance = cv2.pointPolygonTest(mask,tuple(point[0]),1)
    if distance > 0:

and vice-versa (in case if mask is smaller than the frame):

for point in mask:
    distance = cv2.pointPolygonTest(hull,tuple(point[0]),1)
    if distance > 0:

The above will tell if the motion contour extracted is within of area interest.
In addition to the checking if the motion happens withing of area of interest the severity/size of motion can be calculated by calculating the area of the hull via the following:

area += cv2.contourArea(hull)

which then can be compared with the total area:

surface = cv2.contourArea(mask)

The ratio can be converted to a percentage value and thus be used to trigger the recording if the value is above certain threshold:

relative = area * 100.0 / surface

I use ffmpeg for actual recording (it is way more efficient than dumping frames from HD OpenCV capture). I simply launch an ffmpeg subprocess when motion is detected and send a SIGTERM when motion is over:

p=subprocess.Popen(record_cmd,shell=False) # motion start
p.terminate() # motion stop

Note: the ffmpeg will cleanly close the recording if it sent the SIGTERM (opposed to SIGKILL).

For debug and entertainment purposes the image could be displayed via following:

To draw contours and hulls:

for cnt in contours:
    hull = cv2.convexHull(cnt)
    cv2.drawContours(f, [cnt], 0, (0,255,0),1)

Note: The colour is defined by this tuple: (0,255,0)

Then do display the whole thing insert this inside of the while(True):

k = cv2.waitKey(30) & 0xff
if k == 27:

The above is basic idea behind my motion detection scripts. I have omitted a lot of glue logic and arithmetic due to my script is not ready for public display ;).

Extracting GPS data from Viofo A119 and other Novatek powered cameras

The script.
Here it is: nvtk_mp42gpx.py
Alternative version: nvtk_mp42gpx_older.py

What does it do?

This script will attempt to extract GPS data from Novatek MP4 file and output it in GPX format.

Usage: ./nvtk_mp42gpx.py -i<inputfile> -o<outfile> [-f]
        -i input file (will quit if does not exist)
        -o output file (will quit if exists unless overriden)
        -f force (optional, will overwrite output file)

In short: it takes Novatek encoded MP4 file (with embedded GPS data) and extract GPS data in GPX format (as separate file). Note; it does not modify the original MP4 file.

In long:

Continue reading Extracting GPS data from Viofo A119 and other Novatek powered cameras

mini0806 design flaw

Alternative title: why my mini0806 was crashing and stopped working after a light drop.

One of the differences between mini0805 and mini0806 is this heat sink:

It is a bitch to remove, so I gently coerced it with a blade:

After some time of gentle pushing and heating it managed to peel off:

Revealing Samsung ram chip and Ambarella A7LA50 SoC:

Here is why it was not working after slight drop (and crashing before that):

Pulled out pads! Note that the blade I used to remove would cause other side of pads missing if I applied too much pressure.

My theory is following: the heat sink that bridges RAM and SoC is causing stress on these chips. Mostly because it is glued and not mechanically pressed against the chips. This is a very silly design flaw, if for example, only SoC would have the heat sink (like found on IP cameras) then there would not be any stresses. I don’t think RAM needs heat sinking on those things.

I am tempted to take dremel to my other mini and split the heat sink in two…

Reverse engineering Hikvision SADP Tool

I got couple of Hikvision cameras that needed to have their passwords reset.

Instead of reset-to-factory default button these cameras have very elaborate password reset process.

Officially one must download SADP tool, get the serial number off the camera, fetch it to the Hikvision support, then they generate you a reset code that you plug in into the camera.

The unreliable Hikvision support can be bypassed with this tool (more details here).

I feel very dirty because I had to install the SADP in a Windows virtual machine (it does not work under Linux).
Interesting that the tool is build around QT and libpcap so technically it should not be too difficult to port it to Linux.

Looking at traffic captures the tool discovers the camera via multicast (, udp port 37020) with this payload:

<?xml version="1.0" encoding="utf-8"?><Probe><Uuid>13A888A9-F1B1-4020-AE9F-05607682D23B</Uuid><Types>inquiry</Types></Probe>

The camera responds to with this:

<?xml version="1.0" encoding="UTF-8"?>
<SoftwareVersion>V5.2.5build 141201</SoftwareVersion>
<DSPVersion>V5.0, build 140714</DSPVersion>
<BootTime>2016-03-06 09:18:17</BootTime>

This is all nice and easy to replicate, except when discovering that when resetting the password the tool talks to camera directly via ethernet frames:

Reset packet:

12:14:16.063953 52:54:00:db:ae:e4 > XX:XX:XX:XX:XX:XX, ethertype Unknown (0x8033), length 80: 
        0x0000:  2102 0142 0000 173a 0604 0a00 ba54 5254  !..B...:.....TRT
        0x0010:  00db aee4 0a01 0102 XXXX XXXX XXXX 0a01  ........XXXXXX..
        0x0020:  01fb ffff ff00 5252 5364 5264 6572 6439  ......RRSdRderd9
        0x0030:  0000 a100 0000 0000 0000 0000 0000 0000  ................
        0x0040:  0000                                     ..

Response packet:

12:14:16.094857 XX:XX:XX:XX:XX:XX > 52:54:00:db:ae:e4, ethertype Unknown (0x8033), length 260: 
        0x0000:  2101 01f6 0000 173a 0604 0b01 8a3b XXXX  !......:.....;XX
        0x0010:  XXXX XXXX 0a01 01fb ffff ffff ffff 0000  XXXX............
        0x0020:  0000 ffff ff00 4453 2d32 4344 3234 3332  ......DS-2CD2432
        0x0030:  462d 4957 3230 3135 3031 3236 4343 4348  F-IW20150126CCCH
        0x0040:  XXXX XXXX XXXX XXXX XX00 0000 0000 0000  XXXXXXXXX.......
        0x0050:  0000 0000 0000 0000 9812 0000 1f40 0000  .............@..
        0x0060:  0001 0000 0000 5635 2e32 2e35 6275 696c  ......V5.2.5buil
        0x0070:  6420 3134 3132 3031 0000 0000 0000 0000  d.141201........
        0x0080:  0000 0000 0000 0000 0000 0000 0000 0000  ................
        0x0090:  0000 0000 0000 5635 2e30 2c20 6275 696c  ......V5.0,.buil
        0x00a0:  6420 3134 3037 3134 0000 0000 0000 0000  d.140714........
        0x00b0:  0000 0000 0000 0000 0000 0000 0000 0000  ................
        0x00c0:  0000 0000 0000 3230 3136 2d30 332d 3036  ......2016-03-06
        0x00d0:  2030 393a 3138 3a31 3700 0000 0000 0000  .09:18:17.......
        0x00e0:  0000 0000 0000 0000 0000 0000 0000 0000  ................
        0x00f0:  0000 0000 0000                           ......

Now looking further it appears that discovery as well as expected UDP communication there is also ethernet frame type of communication going on in parallel:


12:13:29.539493 52:54:00:db:ae:e4 > ff:ff:ff:ff:ff:ff, ethertype Unknown (0x8033), length 80: 
        0x0000:  2102 0142 0000 1739 0604 0300 80b6 5254  !..B...9......RT
        0x0010:  00db aee4 0a01 0102 ffff ffff ffff 0000  ................
        0x0020:  0000 0000 0000 fe80 0000 0000 0000 889f  ................
        0x0030:  720d 7c8f 8429 0000 0000 0000 0000 0000  r.|..)..........
        0x0040:  0000                                     ..


12:13:29.555356 XX:XX:XX:XX:XX:XX > 52:54:00:db:ae:e4, ethertype Unknown (0x8033), length 416: 
        0x0000:  2101 01f6 0000 1739 0604 0400 8c42 XXXX  !......9.....BXX
        0x0010:  XXXX XXXX 0a01 01fb ffff ffff ffff 0000  XXXX............
        0x0020:  0000 ffff ff00 4453 2d32 4344 3234 3332  ......DS-2CD2432
        0x0030:  462d 4957 3230 3135 3031 3236 4343 4348  F-IW20150126CCCH
        0x0040:  XXXX XXXX XXXX XXXX XX00 0000 0000 0000  XXXXXXXXX.......
        0x0050:  0000 0000 0000 0000 9812 0000 1f40 0000  .............@..
        0x0060:  0001 0000 0000 5635 2e32 2e35 6275 696c  ......V5.2.5buil
        0x0070:  6420 3134 3132 3031 0000 0000 0000 0000  d.141201........
        0x0080:  0000 0000 0000 0000 0000 0000 0000 0000  ................
        0x0090:  0000 0000 0000 5635 2e30 2c20 6275 696c  ......V5.0,.buil
        0x00a0:  6420 3134 3037 3134 0000 0000 0000 0000  d.140714........
        0x00b0:  0000 0000 0000 0000 0000 0000 0000 0000  ................
        0x00c0:  0000 0000 0000 3230 3136 2d30 332d 3036  ......2016-03-06
        0x00d0:  2030 393a 3138 3a31 3700 0000 0000 0000  .09:18:17.......
        0x00e0:  0000 0000 0000 0000 0000 0000 0000 0000  ................
        0x00f0:  0000 0000 0000 029c 5648 0a01 0101 0000  ........VH......
        0x0100:  0000 0000 0000 0000 0000 0000 0000 0000  ................
        0x0110:  0000 0000 0000 0000 0000 0000 0000 0007  ................
        0x0120:  0000 0000 0000 0000 0000 0000 0000 0000  ................
        0x0130:  0000 0000 0000 0000 0000 0000 0000 0050  ...............P
        0x0140:  0000 0000 0000 0000 0000 0000 0000 0000  ................
        0x0150:  0000 0000 0000 0000 0000 0000 0000 0000  ................
        0x0160:  0000 4453 2d32 4344 3234 3332 462d 4957  ..DS-2CD2432F-IW
        0x0170:  0000 0000 0000 0000 0000 0000 0000 0000  ................
        0x0180:  0000 0000 0000 0000 0000 0000 0000 0000  ................
        0x0190:  0000

So theoretically it is possible to create a tool based on the reset password code generator to completely cut out middle man.
This is the way I see it working:

1) Discover camera and get serial number and camera ip
2) Get camera date/time via simple GET to port 80.
3) Generate reset code with camera serial number and date/time
4) send magic packet to reset camera.

I found some examples of the traffic that does not contain fe80 0000…XXXX…0000 bit at the end (looks like previous version of SADP Tool didn’t append that crap). I successfully replayed that packet.
I have noticed that the checksum does not include the source (header) of the packet, so as long as the MAC address matches in the body the header can be spoofed.

I have changed the mac address on VM where SADP Tool was running and looks like4 bytes between Source MAC and Destination MAC in the body changes. As well as 2x 2 bytes surrounding 06040300.

If I increment any number by one and decrement @ 0x0018 the packet gets response. This implies that the check sum is only 2 bytes long.

So far I figured out the check sum for older type of discovery packet (without crap at the end of the packet).

the check sum is located here:
0x0010: 00db aee4 0a01 0102 ffff ffff ffff 0000
In this example it is 0102.
Actually the check sum is 0201 (reversed order).
The check sum algorithm is 16-bit one’s complement.
The trick (which was given away by comparing sequential packets) is to ignore the header, and to reverse order in these two bytes:
0x0000: 2102 0142 0000 1739 0604 0300 80b6 5254
in example above they are check-summed as b680.

Next step is to see if I can apply the same method to the password reset packet….

At this stage I solved the following: discovery via frames, discovery via UDP, generate reset code and reset the camera via frame.

There is a potential problem to get camera time reliably (in case it is not configured in same subnet).

After poking around Sadp.dll I found these interesting XML strings:

<?xml version="1.0" encoding="utf-8"?><Probe><Uuid>%s</Uuid><Types>inquiry</Types></Probe>
<?xml version="1.0" encoding="utf-8"?><Probe><Uuid>%s</Uuid><Types>update</Types><MAC>%s</MAC><Password>%s</Password><IPv4Address>%s</IPv4Address><CommandPort>%d</CommandPort><HttpPort>%d</HttpPort><IPv4SubnetMask>%s</IPv4SubnetMask><IPv4Gateway>%s</IPv4Gateway><IPv6Address>%s</IPv6Address><IPv6Gateway>%s</IPv6Gateway><IPv6MaskLen>%d</IPv6MaskLen><DHCP>%s</DHCP></Probe>
<?xml version="1.0" encoding="utf-8"?><Probe><Uuid>%s</Uuid><MAC>%s</MAC><Types>reset</Types><Code>%s</Code></Probe>
<?xml version="1.0" encoding="utf-8"?><Probe><Uuid>%s</Uuid><MAC>%s</MAC><Types>reset</Types><Code>%s</Code><Password>%s</Password></Probe>
<?xml version="1.0" encoding="utf-8"?><Probe><Uuid>%s</Uuid><MAC>%s</MAC><Types>reset</Types><SyncIPCPassword>true</SyncIPCPassword ><Code>%s</Code><Password>%s</Password></Probe>
<?xml version="1.0" encoding="utf-8"?><Probe><Uuid>%s</Uuid><MAC>%s</MAC><Types>getcode</Types></Probe>
<?xml version="1.0" encoding="utf-8"?><Probe><Uuid>%s</Uuid><MAC>%s</MAC><Types>exchangecode</Types><Code>%s</Code></Probe>
<?xml version="1.0" encoding="utf-8"?><Probe><Uuid>%s</Uuid><MAC>%s</MAC><Types>activate</Types><Password>%s</Password></Probe>
<?xml version="1.0" encoding="utf-8"?><Probe><Uuid>%s</Uuid><MAC>%s</MAC><Types>getencryptstring</Types></Probe>

None of that proved to be useful of extracting local time (except inquiry).

See this post for actual script.