Bosch AI Camera Trainer Released And Tested

By IPVM Team, Published Apr 09, 2019, 10:01am EDT

Bosch is releasing a highly unusual new AI feature - 'Camera Trainer'. Now, coming as a standard feature in Bosch IVA/EVA analytics, one can train the camera to detect custom objects one wants.

image

While detecting a face or a person or a car is increasingly common, Bosch is allowing users to define what objects are most interesting to them - whether it is a drill or a door opening or a ladder or a buggy or whatever might be important to one's camera scene.

On the positive side, this opens up many possibilities. On the negative side, it raises the question of how well this will work and how hard it is to set it up.

IPVM tested the beta version of Camera Trainer, configuring it in multiple scenes, to see how it performed.

*******

** *** *****, ***** Camera ******* ********* ********* as *********, ******** ** to ***** ** ******** objects ***** ***** *** be ************* ** ******** analytics (**** ** ****** being ******* ** ******* left ** * ****) as **** ** ******** on *** ***** ** some *******, **** ** doors ***** ****** ** closed.

*******, ***** *** ***** key ***********:

  • ***** ********* ** ***** subjects:****** ******** *** ***** would ***** ***** *** camera ** ****** ** object **** *** *** exist, **** ** * fold ** * ******'* clothes ********** ** * drill ** * ***** appearing ** ** **** door. ***** ****** ***** not ** ******** ** adding **** ******** *******.
  • ****/*********** **** ****** *******:****** ******* ******** **** object ***********, ****** *****, and **** ****** ******* to ******** ****. **** means **** ******* ******** depth ** *** ***** (moving ****** ** *******) or ******** ********* ****** in ****** ********* ** duplicate ********* ** **** cases. *******/****** ** *********** angled ******* **** ** trained **********.
  • ******* *** ***** ***********:** *** *****, ***** 1 *** *** *****, Camera ******* ****** ** detect ******* ** ***** otherwise ******** ***** **. This ***** ** ******** with ********** ***** (******* or **) ** *** scene, *** **** ** the ********* ****** ********* include ********** **.

****: **** ** * test ** * ********** version ** ****** *******. We **** ** ****** and ****** **** ****** if *********** *********** *** found **** ** **** Bosch's ******** **** ****** Flexidome, ******** ** *****/*** 2019.

************

****** ******* ** ******** to ** ********* ** the *** ** ***** 2019. ** ** ********* on ****, *, *** 7.3 ******, ***** ********:

  • ****** ** ********* ****/****
  • ********* ** ********* ****/**** IC
  • ********* ** ********* *** ultra ****/****
  • *** ** ********* *****
  • *** ** ****** ***** (color ******* ****, *** thermal)

App/Training ********

****** ******* ******** * few ****** **** *** user, ********** ***** *****'* Configuration ******* *******:

  • ******** *******: ***** ******* are ******** ** ******* a *** ****** * specific ****** ** ** trained, **** ** * drill, * ****** ****, a ******, ***. ***** is ** ******** ****** of ******* ******** *** training.
  • ******** *******: ***** ******* are ******** ****** ********, but ******* * *** around * ****** ***** the ****** ************, ** ** ******** on * ******** ***** positive *** ******* ** as ********.
  • ****** ** ********:  *** area ***** ********* ***** place. ** *** *****, ROI *** *** ******** or ******** ******** *** was ****** **** ** narrow ********* ****.

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

HOG **. *** **************

***** ****** **** **** initial ******* ** ****** Trainer ** ***** *** (********* ** ******** *********) *** ****** ***********. However ** *** **** key ***********, ** ** cannot ********* *****, ******* drastically ** ***** ******** (e.g. *** *****) ****** where ***** ** ******* become **** *****, *** must ** ******* ** specific ****** ******, ******* of *******.

***** *** ****-**** ***** to ********* *********** ****** Networks (****) *** ********* in *** ******, ***** should ******* **** **** of ***** ******. ** the ***-****, **** **** plan ** ********* ***-******* models *** ****** ****** classes ** ****** ********.

*** **** ******* ** the *********** ******* ***** approaches, *** ********* ********* ************.

Camera ******* ***********

*****'* ************* ******** ******* regarding ****** ******* ***********, most ************* ****/***********:

****** ******* ********* *** limited ** *** ****. Due ** ***********, ******* appear **** ******* ** the ********** **** **** the ****** *** ***** much **** *****. **** difference ** **** *** pixel ******** ****** *** image ****** ** ******* with ****** *******. **** a ****** ******** ** restricted ********** *** ********* distance, ****** ******* ********* can ** ******* ** compensate *** ***-***** **********. Objects ** *** *** distance ****** ** ******** at ***, ** **** will *** ***** ****** pixel *** *** ****** Trainer ********.

*******, ******* ***** **************, we ***** ****** ******* performed ****** **** ******** with ****** ** ******** changes ** *********** *** size, **** **** *********** than *** ************ ********** Bosch **** ***. ** discuss ***** ****** *****.

Performance ********

**** ********* *****'* ************* guidelines, *** ****** ******* was ****** ******* ** recognize **** ** ***** and **** *** **. In *** ******* *****, the ****** *** ******* to ********* **** ******* cabinets **** ******, **** an ***** ** **** open *** ****. **** that *** ****** ** able ** *********** ******* the **** ******* *** person, ****** *** ******* is ****** ******** ****** him.

Cabinet-Open_Close-Status-Detected-Properly

** ******* **** ***** configuration, *** ****** *** trained **** ****** ** all ******** **** *** closed, ** **** ** negative ******* ** * human ** ***** ** open *** ****** *****.

Training Data (Positive And Negative)

***** *****'* ******* ***** configuration, *** ****** ******* was **** ******* ** count *** **** ******** were **** ****.

Camera Trainer Data Used For Counting

** ******* *******, *** camera *** ******* ** recognize ****** ** * pegboard ****, **** ** alarm ******* **** **** were *******.

*******, **** ***** *** have **** ******, **** other ******* ************ **** as ******, ********** ****** entering *** *****, *** also ***** ***** ** the ****. ** ****** of ********* ***** *** to *** ******* ******* these ***** ******.

Humans Trigger False Detection

**** **** ***** ****** cleared ***** *** ****** left *** *****, ** if *** ****** *** to ****** * ******* number ** *******, *** rule ***** ***** ********.

Perspective *** ****** **** ******

******** ******* **** ** roughly *** **** **** and *********** ** ***** for ****** ******* ** detect ****. *** ***** object ******** *** ******** is *** **** ****** must **, ** *** Trainer's ********* *** **** is ******. **** ***** that **** ****** ** detected **** **** *** side *** ****-**, *** example, ** **** ***** change *** ******'* ****** ratio **********.

******* ** ***** ******, objects **** **** ******** shapes *** ***** **** not ******* ****. *** example, ***** *** ****** could ** ****** ******* to ****** ******* ****** in *** ***** ** view (*****), ****** **** nearer ** ******* **** the ****** ** ******** their *********** ******** ** missed **********.

Perspective Angle Impacts Accuracy

**** ******* ** ****** a ******* ******, *** camera ******** ** ** two ******** ******* **** close ** *** ******, or *** ** ***, since ** *** **** larger **** *** ***** the ****** *** ******* on.

Duplicate Objects Detected At Close Distance

**** **** ***** ********** against ****** ********* **** different ************ *** ********* to * ****** ****** Trainer ****, ** ** may ****** *********** **** looking *** ******** *******. However, ** ********** **** may ** ******* *** trained **** ******** **** an ********** ***** ** distance.

Reduced *********** ** *** *****

****** ******* *********** ** low ***** ********** (~*.***) was *******, ***** ******* objects ** *** ******* to ***** ** **** it ***** ******** ***** on ** **** *****. This ***** ** ******** using ******** ************, *** no ********* ****** ******* integrated **.

Low-Light-Drastically-Reduces-Performance-(~0 5-Lux)

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

***** ****** ******* ** integrated ****** ***** ******** analytics, **** ******** ***** used ** ******* ** standard **** ***** (****** detected, *********, ****** *******, etc.). ******* ** ****, Camera ******* ** *********** compatible **** *** *** that ************* *********, ********* *****, *******, and *********.

*******, **** **** ***** may *** ** *********. For *******, *** "*********" rule ** ********* ** "People ********" ** ******* Security ******, ****** ** does ********* ******* *** number ** ******* *******:

VMS Integration Same As IVA

Versions ****

*** ********* ******** **** used ****** *******.

  • ***** ************* *******:*.**.****.* - **** **** this ******* ** * pre-release ****, *** ********* for ****** ***. *** non-beta ******* ** ******** in *** **** *** weeks.
  • ***** ** ********* **** VR:*.**.****

Comments (11)

Interesting concept. Nice test.

Agree
Disagree
Informative
Unhelpful
Funny

The customization of scenes and processing models based thereon is probably something that will be offered as a service from vendors (herein "Bosch") working in coordination with the customer.  Bosch will have more samples available to it of different things and can produce a superior model.  Rather than pay customer company employees to try and master models and risk the inherent errors or just the time consumption, it will be more efficient for Bosch to offer as a service the customization of camera scene where Bosch's computers do the heavy lifting of training and then download the final model to the customer's camera.  With super charged AI machines running in the $100,000s (see NVidia's DGX-1 at $150,000 and DGX-2 at $400,000) and the time and power it takes to create good models, it seems counterproductive to expect customers to fabricate good models when Bosch has the tools to accomplish the job.  I think it will be more efficient for Bosch to sell configuration as a service.

Agree: 3
Disagree
Informative: 1
Unhelpful
Funny

call me when it can detect a baggage left alone at the checkin area at an airport.

Agree
Disagree
Informative
Unhelpful
Funny

The Object left behind analytic has been in place with Bosch for a while. The Camera trainer analytic is designed to provide a more customized solution for a customer. The customer simply gives the cameras positive and negative examples of what it is being trained to look for. these examples come straight from the scene via the camera. The more examples given the more accurate the function is. I have trained it to recognize a small banner and within 10 min it was recognizing the banner and ignoring other things in the scene.

Agree
Disagree
Informative
Unhelpful
Funny

Bosch has long-term plans to implement Convolution Neural Networks (CNNs) for detection in the future, which should improve upon some of these issues.

If Bosch implements CNNs for detection, this will effectively prevent most customers from training it themselves since training CNNs requires thousands of labeled sample images. The current HOG implementation by contrast is incredibly simple to train by a customer. But it is nowhere near as accurate as CNNs.

Agree: 2
Disagree
Informative: 2
Unhelpful
Funny

Is there something you have to enable in the camera or in the Configuration Manager software for the Camera Trainer tab to show up in Configuration Manager?

I have installed Configuration Manager 6.20 and have firmware 6.60 using the same camera used in one of the videos above but I don't get the tab.

Agree
Disagree
Informative
Unhelpful
Funny
Here is a link to the Data sheet on the Bosch Wed Site: Intelligent Video Analytics 7.10 In addition the part numbers for the Licenses are in the Data sheet, and there is no charge for the licenses.
Agree
Disagree
Informative
Unhelpful
Funny

UM#3, the use of camera trainer requires the purchase of a license, once the license is purchased the camera trainer tab will be available in configuration manager.

Agree
Disagree
Informative: 1
Unhelpful
Funny

The Camera Trainer license does not have a cost associated with it. It is a free license.

The integrator must complete a free online training about 45 mi to an hour long to have the free license released to them.

Bosch does this to ensure the integrator has the basic working knowledge needed to set up the camera trainer.

Agree
Disagree
Informative: 3
Unhelpful
Funny

This is the 1st time I have heard of the HOG and CNN AI - is this a standard for all MFGS?
HOG came 1st I guess and CNN is the better solution? not sure I understand the differences.
But I am loving the AI so far.

Agree
Disagree
Informative
Unhelpful
Funny

We have a good primer report here: IPVM Deep Learning Tutorial

HOG and CNNs are both building blocks for video analytics. HOG is an older generation, hardcoded computer vision algorithm, whereas a CNN (convolutional neural network) is a deep learning system which is what many modern AI video surveillance analytic systems are based on.

Deep learning systems are trained with datasets and learn what objects are based on that training. HOG/computer vision systems are preprogrammed with parameters of what shape/motion a (person/dog/face/bag) looks like.

We have tested different CNNs/hardware platforms, and also had a Deep Learning Class earlier in 2019:

Agree
Disagree
Informative
Unhelpful
Funny
Read this IPVM report for free.

This article is part of IPVM's 6,956 reports, 927 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