7 ms·

The sensor technology is a big factor here. You can use LIDAR for perception without needing a neural net. On the other hand with radar + ultrasound + vision yo

by ummonk 8y ago

The sensor technology is a big factor here. You can use LIDAR for perception without needing a neural net. On the other hand with radar + ultrasound + vision you'll need a neural net to try to detect things like obstacles, since the raw data from those sensors doesn't directly tell you whether there is an obstacle in the way.

ebg13 8y ago
Everything you said is very wrong. LIDAR produces line of sight depth information. Stereo cameras produce line of sight depth information. Radar produces line of sight depth information. Ultrasound produces line of sight depth information. The only differences are how, how fast, and how detailed. The biggest difference between LIDAR and stereo vision is that LIDAR doesn't depend on illumination because it uses its own.
- gamblor956 8y ago
  Stereo cameras produce line of sight depth information That is wrong. Stereo cameras provide data which, after significant processing, can result in line-of-sight depth information. With radar (including LIDAR) and ultrasound, the depth information is essentially "free" since very little processing is required to extract the depth information (i.e., the information returned is the depth). So the biggest difference between stereovision and everything else is that stereovision requires additional processing to extract the depth information.
  - colordrops 8y ago
    Aren't the algorithms for extracting depth from stereo imagery already solved and thus also essentially "free"?
    - codelord 8y ago
      Depth estimation is not a solved problem. Why do you think SDC companies pay so much to buy Lidar if depth estimation was solved?
      - unityByFreedom 8y ago
        We need to define solved. Used in the Musk sense, it is solved because it's possible. Used in the academic sense, it's not as reliable as depth from LiDAR.
        rohit2412 8y ago
        This is the Crux of the problem. Elon musk and Tesla are happy to use words with misleading meaning deliberately. Autopilot, full self driving, "solved", safer than a human driver etc.
        natch 8y ago
        There's nothing wrong with using words. The real key is to understand carefully the context in which the words were used. The "misleading" accusation is just a projection people like you are imagining after cherry picking words without paying attention to the context. The words you are missing are things like: "will be," "in the future," "it is possible," "assuming regulatory approval," etc.
      - pbreit 8y ago
        Because they bet wrong. It happens.
        shaklee3 8y ago
        Did that Tesla that smashed into the side of a semi truck bet correctly? https://www.google.com/amp/s/electrek.co/2016/07/01/understanding-fatal-tesla-accident-autopilot-nhtsa-probe/amp/ https://www.google.com/amp/s/electrek.co/2016/07/01/understa...
        areoform 8y ago
        ‘Did that Tesla that smashed into the side of a semi truck bet correctly?’ That’s needlessly antagonistic and not cool.
    - gamblor956 8y ago
      No, it's like this: Radar, LIDAR, ultrasound: returning packets have X value, where X is run through a single algorithm to return depth. In the case of ultrasound and radar, the strength of the return packet is generally directly proportional to depth (algorithmically). LIDAR systems might require an additional processing step depending on the type of LIDAR system used, since many systems use shifted-arrays or rotating frequencies. Stereovision: returns two matrices of data (i.e., each image). Each matrix is run through edge-detection and object-detection algorithms, and the resultant data is then run through more algorithms to match and compare objects between the two images. Finally, once objects have been matched up, a parallax algorithm determines their 3d positioning. While this can all be done in hardware (i.e., ASICS), it's still multiple additional levels of processing that must be done before you can even get the data to the decision-making part of the system. Stereovision is thus always inherently laggier than a radar-based system when it comes to determining depth.
      - ebg13 8y ago
        > Each matrix is run through edge-detection and object-detection algorithms, and the resultant data is then run through more algorithms to match and compare objects between the two images. Finally, once objects have been matched up, a parallax algorithm determines their 3d positioning. Where is your information coming from, and can you point me to a stereo algorithm that uses object detection?
        jeffreygoesto 8y ago
        Right, almost all stereo depth calculation is matching without object detection. Maybe doing both simultaneously counts? ;-) https://arxiv.org/abs/1902.09738 https://arxiv.org/abs/1902.09738
      - pbreit 8y ago
        Sure it’s laggier but with dedicated silicon that’s a non-issue. Fact is vision is more than adequate. And lidar is literally blind.
        8y ago
  - pbreit 8y ago
    Sure, but if you’ve got the processing it’s pretty easy.
areoform 8y ago
That's a very common perception, but it's not true! Kalman filters have for years been used to get usable data out of these and multiple studies have shown that ultrasound can be used as a standalone in low-speed scenarios. Design and experimental study of an ultrasonic sensor system for lateral collision avoidance at low speeds - https://ieeexplore.ieee.org/abstract/document/1336460 https://ieeexplore.ieee.org/abstract/document/1336460 Environmental perception and multi-sensor data fusion for off-road autonomous vehicles - https://ieeexplore.ieee.org/abstract/document/1520113 https://ieeexplore.ieee.org/abstract/document/1520113 (this one uses laser range finders not LIDAR with some vision) "Automotive radar the key technology for autonomous driving: From detection and ranging to environmental understanding" - https://ieeexplore.ieee.org/abstract/document/7485214 https://ieeexplore.ieee.org/abstract/document/7485214 Real-time SLAM has also been done with vision + inertial navigation for UAVs; A synchronized visual-inertial sensor system with FPGA pre-processing for accurate real-time SLAM - https://ieeexplore.ieee.org/abstract/document/6906892 https://ieeexplore.ieee.org/abstract/document/6906892 - https://www.research-collection.ethz.ch/bitstream/handle/20.500.11850/81914/eth-8003-01.pdf https://www.research-collection.ethz.ch/bitstream/handle/20.... Some people have proposed laser range finders with ultrasound (sonar); 'This article addresses the problem of performing SLAM under reduced visibility conditions by proposing a sensor fusion layer which takes advantage from complementary characteristics between a laser range finder (LRF) and an array of sonars. This sensor fusion layer is ultimately used with a state-of-the-art SLAM technique to be resilient in scenarios where visibility cannot be assumed at all times.' A Sensor Fusion Layer to Cope with Reduced Visibility in SLAM - https://link.springer.com/article/10.1007/s10846-015-0180-8 https://link.springer.com/article/10.1007/s10846-015-0180-8 - http://eprints.lincoln.ac.uk/16820/7/__ddat02_staffhome_jpartridge_A%20sensor%20fusion%20layer.%20Santos.pdf http://eprints.lincoln.ac.uk/16820/7/__ddat02_staffhome_jpar... There are enough datapoints between vision, ultrasound and radar for old school algorithms to do real time SLAM and make cost maps. So the problem and solution definitely isn't MOAR SENSARS.
- gugagore 8y ago
  Kalman filters over what state space? Localization isn't really the challenging aspect of self-driving cars.
pbreit 8y ago
That’s baloney and obviously wrong based on eyesight.

by ummonk 8y ago

ebg13 8y ago

Everything you said is very wrong. LIDAR produces line of sight depth information. Stereo cameras produce line of sight depth information. Radar produces line of sight depth information. Ultrasound produces line of sight depth information. The only differences are how, how fast, and how detailed. The biggest difference between LIDAR and stereo vision is that LIDAR doesn't depend on illumination because it uses its own.

gamblor956 8y ago
Stereo cameras produce line of sight depth information That is wrong. Stereo cameras provide data which, after significant processing, can result in line-of-sight depth information. With radar (including LIDAR) and ultrasound, the depth information is essentially "free" since very little processing is required to extract the depth information (i.e., the information returned is the depth). So the biggest difference between stereovision and everything else is that stereovision requires additional processing to extract the depth information.
- colordrops 8y ago
  Aren't the algorithms for extracting depth from stereo imagery already solved and thus also essentially "free"?
  - codelord 8y ago
    Depth estimation is not a solved problem. Why do you think SDC companies pay so much to buy Lidar if depth estimation was solved?
    - unityByFreedom 8y ago
      We need to define solved. Used in the Musk sense, it is solved because it's possible. Used in the academic sense, it's not as reliable as depth from LiDAR.
      - rohit2412 8y ago
        This is the Crux of the problem. Elon musk and Tesla are happy to use words with misleading meaning deliberately. Autopilot, full self driving, "solved", safer than a human driver etc.
        natch 8y ago
        There's nothing wrong with using words. The real key is to understand carefully the context in which the words were used. The "misleading" accusation is just a projection people like you are imagining after cherry picking words without paying attention to the context. The words you are missing are things like: "will be," "in the future," "it is possible," "assuming regulatory approval," etc.
    - pbreit 8y ago
      Because they bet wrong. It happens.
      - shaklee3 8y ago
        Did that Tesla that smashed into the side of a semi truck bet correctly? https://www.google.com/amp/s/electrek.co/2016/07/01/understanding-fatal-tesla-accident-autopilot-nhtsa-probe/amp/ https://www.google.com/amp/s/electrek.co/2016/07/01/understa...
        areoform 8y ago
        ‘Did that Tesla that smashed into the side of a semi truck bet correctly?’ That’s needlessly antagonistic and not cool.
  - gamblor956 8y ago
    No, it's like this: Radar, LIDAR, ultrasound: returning packets have X value, where X is run through a single algorithm to return depth. In the case of ultrasound and radar, the strength of the return packet is generally directly proportional to depth (algorithmically). LIDAR systems might require an additional processing step depending on the type of LIDAR system used, since many systems use shifted-arrays or rotating frequencies. Stereovision: returns two matrices of data (i.e., each image). Each matrix is run through edge-detection and object-detection algorithms, and the resultant data is then run through more algorithms to match and compare objects between the two images. Finally, once objects have been matched up, a parallax algorithm determines their 3d positioning. While this can all be done in hardware (i.e., ASICS), it's still multiple additional levels of processing that must be done before you can even get the data to the decision-making part of the system. Stereovision is thus always inherently laggier than a radar-based system when it comes to determining depth.
    - ebg13 8y ago
      > Each matrix is run through edge-detection and object-detection algorithms, and the resultant data is then run through more algorithms to match and compare objects between the two images. Finally, once objects have been matched up, a parallax algorithm determines their 3d positioning. Where is your information coming from, and can you point me to a stereo algorithm that uses object detection?
      - jeffreygoesto 8y ago
        Right, almost all stereo depth calculation is matching without object detection. Maybe doing both simultaneously counts? ;-) https://arxiv.org/abs/1902.09738 https://arxiv.org/abs/1902.09738
    - pbreit 8y ago
      Sure it’s laggier but with dedicated silicon that’s a non-issue. Fact is vision is more than adequate. And lidar is literally blind.
      - 8y ago
- pbreit 8y ago
  Sure, but if you’ve got the processing it’s pretty easy.

areoform 8y ago

That's a very common perception, but it's not true! Kalman filters have for years been used to get usable data out of these and multiple studies have shown that ultrasound can be used as a standalone in low-speed scenarios. Design and experimental study of an ultrasonic sensor system for lateral collision avoidance at low speeds - https://ieeexplore.ieee.org/abstract/document/1336460 https://ieeexplore.ieee.org/abstract/document/1336460 Environmental perception and multi-sensor data fusion for off-road autonomous vehicles - https://ieeexplore.ieee.org/abstract/document/1520113 https://ieeexplore.ieee.org/abstract/document/1520113 (this one uses laser range finders not LIDAR with some vision) "Automotive radar the key technology for autonomous driving: From detection and ranging to environmental understanding" - https://ieeexplore.ieee.org/abstract/document/7485214 https://ieeexplore.ieee.org/abstract/document/7485214 Real-time SLAM has also been done with vision + inertial navigation for UAVs; A synchronized visual-inertial sensor system with FPGA pre-processing for accurate real-time SLAM - https://ieeexplore.ieee.org/abstract/document/6906892 https://ieeexplore.ieee.org/abstract/document/6906892 - https://www.research-collection.ethz.ch/bitstream/handle/20.500.11850/81914/eth-8003-01.pdf https://www.research-collection.ethz.ch/bitstream/handle/20.... Some people have proposed laser range finders with ultrasound (sonar); 'This article addresses the problem of performing SLAM under reduced visibility conditions by proposing a sensor fusion layer which takes advantage from complementary characteristics between a laser range finder (LRF) and an array of sonars. This sensor fusion layer is ultimately used with a state-of-the-art SLAM technique to be resilient in scenarios where visibility cannot be assumed at all times.' A Sensor Fusion Layer to Cope with Reduced Visibility in SLAM - https://link.springer.com/article/10.1007/s10846-015-0180-8 https://link.springer.com/article/10.1007/s10846-015-0180-8 - http://eprints.lincoln.ac.uk/16820/7/__ddat02_staffhome_jpartridge_A%20sensor%20fusion%20layer.%20Santos.pdf http://eprints.lincoln.ac.uk/16820/7/__ddat02_staffhome_jpar... There are enough datapoints between vision, ultrasound and radar for old school algorithms to do real time SLAM and make cost maps. So the problem and solution definitely isn't MOAR SENSARS.

gugagore 8y ago
Kalman filters over what state space? Localization isn't really the challenging aspect of self-driving cars.

pbreit 8y ago

That’s baloney and obviously wrong based on eyesight.