Subj : Re: Collision detection again To : rec.sport.rowing From : Kit Davies Date : Fri Aug 14 2020 04:38 pm On 14/08/2020 07:05, Peter wrote: > On Friday, 14 August 2020 at 00:57:00 UTC+1, carl wrote: >> On 13/08/2020 14:03, Andy McKenzie wrote: >>> On Thursday, 13 August 2020 at 02:10:03 UTC+1, A. Dumas wrote: >>>> carl wrote: >>>>> On 12/08/2020 21:51, marko....@gmail.com wrote: >>>>>> I do think this problem is solvable with current computer vision >>>>>> technology though, the problem is getting enough data and the right expertise on it. >>>>>> >>>>> >>>>> The fundamental test would appear to be that a) an object remains on the >>>>> same heading WRT your own path and b) that the distance is closing. >>>>> That seems rather easy. >>>> You'd think. >>>> https://www.heise.de/hintergrund/Pixelmuster-irritieren-die-KI-autonomer-Fahrze uge-4852995.html?seite=all >>>> >>>> Research basis for that article, linked from there but in English: >>>> https://arxiv.org/abs/1707.08945 >>> >>> Several things about rowing make an Image recognition approach for obstacle detection an interesting prospect. Sculling hazards manifest in a relatively leisurely manner and at relatively slow speed. We don't generally have accidents because someone darts out of a tributary, we have them because we fail to see an obstacle, often because we assume that there shouldn't be an obstacle there (the classic excuse of 'I was on my side of the river' or 'I hadn't realised I'd drifted over the centre line'). That gives the AI time to 'think'. The camera can look in quite a defined zone as at least on inland waters the view can be limited to a strip of water dead ahead. There should be a low cost to false positives. I imagine a system which has an audio warning 'Ahead sculler'. - that prompts me to look around, which I probably should have been doing anyway. The 'might breed complacency' argument has to be balanced against the material and, potentially, human cost of collisions. On my stretch of river we have rowers training hard or out for a paddle, canoes, paddleboards, rubber rings, buoys, barges, gin palaces and swimmers in every variety from 'long distance with neon hat and float' to child without hi vis. Ironically enough the last damage collision we had was with a 80 foot barge. The barge was on the wrong side of the river, the sculler saw it but misinterpreted which way it was going . To be fair it was weirdly symetrical from bow or stern. Still - a collision warning would have probably avoided a -u1000 repair bill. >>> >>> Given the ready availability of image recognition libraries that will run on hardware like a Raspberry PI this doesn't seem like an insurmountable technical challenge. It would seem like a good university MSc project >>> >> With respect to the brains here more expert in feature recognition and >> detection: why does the potential collision avoidance system need to >> "recognise" anything? All you need to know is that you are closing on a >> collision path with a "something" that is large enough to be detected. >> Doesn't matter what it is, since it is never good to bump into anything, >> be it log, sculler, eight, barge, bridge, the Black Buoy at Putney, the >> Eddystone Lighthouse, whatever. I do favour the KISS principle - until >> proved mistaken, of course. >> Cheers - >> Carl >> >> -- >> Carl Douglas Racing Shells - >> Fine Small-Boats/AeRoWing Low-drag Riggers/Advanced Accessories >> Write: Harris Boatyard, Laleham Reach, Chertsey KT16 8RP, UK >> Find: tinyurl.com/2tqujf >> Email: ca...@carldouglasrowing.com Tel: +44(0)1932-570946 Fax: -563682 >> URLs: carldouglasrowing.com & now on Facebook @ CarlDouglasRacingShells >> >> >> --- >> This email has been checked for viruses by AVG. >> https://www.avg.com > > Just 'deciding' if the way ahead is clear of obstruction is a hugely complex undertaking for a purely optical system. Despite the massive investment in automobile automation such systems are dependant on good road-markings, maps and GPS just to keep the vehicle true. Distance measures via just optics either requires comparisons of previous frames related to speed to determine relative increases in size of possible objects so it first has to determine what an object is (out of a bunch of pixels). Factor in things that bob about from a platform that oscillates and a lot more processing is needed. That leads to ever increasing processor speeds and power requirements and weight on-board. Shadows, bends, varying light conditions and low morning/evening sun all conspire to make it more difficult. False alarms lead to users ignoring them or 'reasonable' efficiency leads to complacency and false trust. Even in ideal motorway conditions a Tesla will throw ocassional phantom braking episodes and if one 'plays' with the automation on A-roads then you'ld better have fast reflexes for when it gets it wrong and be prepared for the ire of drivers behind when it slows to a crawl after misidentifying the speed of the next bend or a sudden patch of shadow. > A simple camera and screen would be a better way forward but even then you need a big enough screen that can be seen in varying light conditions by a camera that can cope with low light or low sun. > pgk > While all true, you're talking about a system with much higher risk profile, given the ubiquity of vehicles and impact (literally and figuratively) of things going wrong. These are also aimed at autonomous operation, which imposes a far tighter degree of tolerance. Object detection is bread-and-butter for modern systems, even in varying light levels, and can be made to run efficiently on constrained devices. The specifics for rowing is identifying how the object is moving in relation to the background, but again, that isn't any more than analysing the difference between images. The main difficulty AISI is in gathering a suitably large training dataset. Eg it will need to experience objects large and small, on wide rivers and narrow. We are helped by the fact that in virtually all cases, there would be a defined path ahead, ie the river, and a reference guide, ie the bow canvas. Incursions into that become your "obstacles". I'm optimistic such a device is achievable. Kit --- SBBSecho 3.06-Win32 * Origin: SportNet Gateway Site (24:150/2) .