Bike-cameras

Incidents

Trouble viewing the videos? Wikipedia has useful instructions. (Wikipedia page `Media help (Ogg)' retrieved Mon, 30 Apr 2007 15:19:28 +0100.)

Distance calibration

Calibration video giving distances in metres as 3.6Mb theora; 4Mb mpeg1.

Photos and demo

The photos show the cameras attached to the bike.

The two videos (the .ogg files) show my journey from Gilbert Road to the Parker's Piece corner at the end of Regent Terrace, in the light, and back again, in the dark. Nothing very unusual happens in these videos, so don't bother watching them for the excitement.

Commentary

The chief difficulty is that the field of view is rather narrow. Many events of interest may occur outside it, especially hazards from the sides. It also gives the impression that things (particularly, vehicles) are much closer than they really are.

The strange jelly-like effect (most visible on the front camera when going over bumps) appears to be caused by the camera not capturing the whole frame at a single instant: so, as the camera moves up or down, objects are vertically squashed or stretched. I was able to reduce this problem from initially unmanageable distortion by greatly increasing the rigidity of the front camera mounting.

At night the imagery is very poor in all but the best-lit streets and my own front light is evidently not strong enough to help this much.

The audio track is nearly useless. In the example videos there are a couple of occasions where I shout warnings to other road users, which are barely audible let alone intelligible. In contrast, brake squeal and road noise are very loud.

Using the camera to record registration plate numbers will be difficult unless the lighting conditions are good and the distances small, and because the audio track is so useless merely calling out the number won't help. I'll have to keep to pencil and paper.

Technical details

The cameras produce .avi files containing a pile of individual .jpeg frames and an uncompressed 8-bit 8khz u-law soundtrack. I have a script which postprocesses a pair of .avi's into a single ~220kby/s .ogg (vorbis audio plus theora video) stream.

The script (written in bash) uses:

• ffmpeg to deconstruct these files;
• djpeg to decompress the jpegs into ppms;
• sox to do audio conversions;
• a crude .c program to determine the time offset between the two video streams by looking for the offset at which the (starts of) the streams correlate most crossly;
• netpbm to tile the two frames, and optionally add pixellation (which sometimes seems appropriate for publication) and optionally to shrink each video frame if I want a small preview.
• ppm2theora-derivative which is a derivative of the libtheora example_encoder.c, modified by me to take as input a stream of maxval-255 raw ppms instead of MPEG4YUV.

This software is available for download. It is in source form only and not very well documented so you will need to be a reasonably competent user of the Unix command line to make use of it. The dependencies are all satisfiable using packages only from Debian etch main.

Campaign for a non-browser-specific WWW

Copyright/acknowledgements.