Bosch Person / Car / Bike / Truck Analytics Tested

By Ethan Ace, Published Jun 05, 2015, 12:00am EDT

Automatically identify what objects are persons vs cars vs bikes vs trucks. That is the new feature and claim from Bosch in their newest IVA analytics (6.10 version). 

We first tested Bosch IVA here. With this new feature, we wanted to see how accurately this worked.

This video shows the basic setup process we used before testing:

We tested these analytics on multiple cameras in a busy outdoor scene, to answer these questions: 

  • Are cars, trucks, people, and bikes reliably detected and classified?
  • What objects are missed most often?
  • How do busy scenes impact performance?
  • Are VMSes able to receive these events distinctly?
  • *****'* *** ****** ************** was **** ******** **** cars *** ******. ** low ******** ******, ********* and ******* ************** ** these ******* ****** * ~190' field ** **** *** ********, missing *** *******.

    ********* ** ****** *** less ********, **** ******** cars ******* ******** ********** as * *****, *** some ******* ****** ********** as ****.

    ******** **** ********** ***** reliably, **** ****** ******* **** strollers **  *** **** cars ******** ** *****. Actual ******** **** ************ missed ** ********** ** another **** ** ******.

    *** ******** *********** *** object **************, ********* ***** to ******* ******** **********/******** lines, ** **** ** one ** **** ****** for ****** *********. *********** was ****, ** ***** measurements *** *****, *** roughly ******** ** *** size/distance ***********.

    ***** *** ****** ********** with ******* ******** ****** and ********* ********, **** distinct ***** ******** ********** (e.g., ****** ********* **. vehicle *********, ***.).

    ***************/******

    **** ***** ***** ***** object ************** *********, *****'* IVA ********* **** * step ****** ** ******** to ****** *** ******** providers **** ********/******* *** AgentVI, ***** **** ******* these ************ *** **** time.

    *******, ***** ****** ** careful ** ***********, *** limit ***** ***** ********* are ********, **** **** activity ****** ********** ******* objects ** ***** *** be *************. ************, ** recommend ******** ************ ** people **. ******** (**** cars *** ******), ** proper ********* ** ******** performed *****.

    Firmware ************

    ***** ************ ** *** are ********* ******** ** firmware *.**, ********* *** most ******* ** ******* in *** ***** ****. Firmware ** ********* *** download **** *** ***** Security *******.

    VMS ***********

    *** ******* ****** ******** events ** ******* ******** Center *** **** *** rule ******* ** *** camera. **** ** ***** in *** ***** ****, here, **** ****, ******, and ******* "****** *** object" ***** ***** **********.

     

    ********* ******** ******** ** to ** ******** ***** analytics ****** **** ***** IVA *******, ****** **** names *** *** ****** from *** ****** *** must ** ********** **********.

    ******** ******* ****** *** Exacq *** *** ******* these ******** ******, **** basic ****** *********.

    Calibration *******

    *********** ******** ***** ** mark ***** *********, *******, and/or ****** ** *** camera's ***** ** ****. We ********* *********** ***** Google **** ************ *** angle ********* *** ***** it ** ** ****** accurate, ****** * *** feet ** ****** ************.

    **** ******* ** ******** in *** ***** *****:

    Detection ***********

    **** ****** ******** ** the *****, ****, ******, and ****** *** ********** correctly >**% ** *** time. *** *******, ****** flags *** **** *** people *** ** **** correctly ******** ** *** respective ******* ** **** image:

    ********* *** ******** *** people, **** **** ******** objects ** *** *****.

    **** *** ****** **** correctly ********** ** *** majority ** *****, ****** there *** **** ******* with ******, ********* ** size (*****):

    *******, **** * **** (bicicyle ** *********) *** present, ** *** ********* classified ** *** *****, such ** *** *** seen ****:

    Misclassification ******

    *******, ****************** **** ****** for *****, *** ** a ****** ******, ******.

    ******* ******* ** "******"

    **** ********, ****************** **** caused ** ******** ******* grouped ******** **** ***** was **** ******** ** the *****, ********* ** a ****** ****** ****** to *** ******. **** was **** ***** *** to * ***** ** cars ********* ** * truck, **** ****:

    ** *** ************* *******, the *********** ******* **** and ****** ** ****** most *****. *********, ***** want ** **** ******* a ******* ** *******, regardless ** *** ****. However, ** **** **** specialized ************ (*.*., ******* or ******* ***** ****** are **********, ****-**** *********, etc.), ********* ** *** distinct ******* ** *********.

    ***** ******* ********* ** Bikes

    ***** **** *** **** common *********** ****** ****, with *********** ***** ***** object **** ********* ** a **** ** *** camera ** *** ***** or *******. 

    *** *******, * ***** subject **** * ********:

    ** ******* * ******:

    **** ** **** ** these *****, *** ****** is *******, ** * similar ******* ** ******* riding * ****, ****** the ****** *** *** misclassification.

    *******, ** *** ***** case, *********, * *** was ********** ** * bicycle ** ****:

     

    Cameras **** 

    *** ******* **** **** for **** ****:

    • ***** ***-*****: *.**.****
    • ***** ***-****: *.**.****

    ******* ******** ****** *.* SR9 *** **** *** recording.

Comments (14)

I'm curious how it would work with a different FOV. Two things I notice:

1) The camera seems relatively low, it looks like maybe 8-9' high. I would expect improved results from a higher vantage point with a little more downward tilt of the camera

2) The FOV seems exceptionally wide and you appear to be getting barrel distortion from the lens. Looks like a ~3mm lens. I'd suggest trying something no less than 4mm to get a flatter image and to keep the overall horizon line and geometry more squared up.

What was the logic in using this FOV? Was it more or less random in the sense you picked a "typical" scene, or did it conform to some recommendations from Bosch?

Are the analytics included with the camera purchase similar to the way Samsung bundles them in with the WiseNetIII. Overall they appear to work fairly well. It will be interesting to see how they perform over a longer period of time. I have had issues with "in camera" analytics where they just stop sending alerts and the camera needs to be rebooted.

They're included on all 7000/8000 series cameras at no charge (5MP starlight, 1080p HDR, 4K, etc.), but not on the 5000 series.

I think that Bosch is trying to do too much.

Avigilon approach of people/cars is much more reliable as I don't see any tech today that can really detect a the difference between a person and a bicycle.

Its better to do less in a good way than to do more poorly.

Finally, when a bike (bicicyle or motorbike) was present, it was correctly classified in all cases...

However, misclassifications were common for bikes...

Meaning other types were commonly misclassified as bikes, since no bikes were actually misclassified?

Correct.

Ethan could you post some video clips from the intersection shot showing objects being detected? Also could you show images/video of the calibration setup for the intersection?

Just following up with the request for some video clips from the street shots.

I would really like to know detection ranges in day and night scenarios, with the lens length or angle used. That to me are the most important numbers along with reliability. Getting an idea of reliable ranges gives you an idea of how many cameras are needed for coverage.

In this test we were trying to get an idea of how well classification worked, not intending this to be a redo of their IVA detection performance. We tested that here (and in the rain). That being said, in this scene, I'd say detection of humans was reliable out to about 150' using the Bosch 5MP, or about a 285' HFOV, something like this:

I know I asked already asked but could you please post some video so we can see the detection range.

The issue with video clips here is that Bosch doesn't embed bounding boxes in the video (like Avigilon no longer does), and to my knowledge, no one has integrated the bounding boxes via metadata other than their own VMS, which we don't use. So we can see them live using Configuration Manager, but Genetec (which fully integrates the separate rules, sans bounding boxes) doesn't display the box.

So I can show you clips and you can see what is in the scene, but it's not going to show you bounding boxes. It's not going to show you events as they happen, either. Video associated with each event can be exported, but multiple events can't be exported with one stretch of video.

I will look through what we have saved as far as config/screen captures of Configuration Manager, as that's the only thing that'll show bounding boxes and path.

The cameras tested here don't support The Bosch CPP6 platform on which this IVA version is supported, therefore tests are i"not really relevant. Re-test on the Starlight 8000 and would be interesting to see results.

1. We used a Starlight 8000, the NBN-80052, as we mentioned in the report. The images in the report are taken from that camera.

2. As per Bosch's documentation IVA 6.10 is supported on more than CPP6.

IP cameras from Bosch are grouped by their common
product platform (CPP) generation. IVA 6.10 is
available on CPP4 and CPP6 based IP cameras

The NBN-932 (now numbered NBN-71027) is CPP4, and supported.

Read this IPVM report for free.

This article is part of IPVM's 6,807 reports, 913 tests and is only available to members. To get a one-time preview of our work, enter your work email to access the full article.

Already a member? Login here | Join now
Loading Related Reports