1. Drones - A New Dimension for Big Data
无人机 – 海量数据的新领域
Yan Ke, PhD 柯严
Chief Software Development Officer
EHang 亿航
Claire Fang, PhD 方芳
VP of Product Development
EHang 亿航
Strata + Hadoop, Beijing 2016
11. Satellites
$50M - $400M to build and launch
Millions per year to operate
Hundreds of Millions
Planes
$300K+ to buy
$100K per year to hire a pilot
Hundreds of Thousands
Drones
$1000 - $10,000 to buy
Autonomous
Thousands
Cost of Data Acquisition
20. Regulatory Support (United States example)
• Manned aircraft pilot’s license
• Section 333 exemption
• Visual line of sight
• Daylight operation
• Can’t fly over people
• One drone per pilot
• Max speed: 100 mph (87 knots).
• Max altitude: 500 feet AGL
Section 333 Exemption Part 107
6 months
Currently application process
5538
total exemptions granted
as of Aug 3rd, 2016
22. DroneDeploy
Automated Flight
Automates the flight and captures all the
aerial data needed.
Cloud-Based Map Creation
Automatically create high-resolution 2D/3D
maps with survey-level accuracy.
25. Types of data transmitted and processed
Real-time control and flight data
– GPS, IMU, Ultrasound
Kb/s
Real-time image, video & thermal data Mb/s
Radar/LIDAR/3D scanners Mb/s
26. Data transmission latency
Satellite 3G Wireless 4G Wireless DSL Cable
750-2000 ms
100-350 ms
50-150 ms
50-125 ms
25-75 ms 20 ms
Gaming VR
minimum acceptable
It’s my pleasure to have been invited here to give a talk about the intersection of two of the most exciting areas in the industry – drones and big data. Not only does the average person have the ability to easily collect massive amounts of data, she now has access to relatively mature platforms and algorithms to be able to process and analyze all of that data. While large corporations and governments have had access to satellites and planes for their remote sensing needs, for the first time in history, small business owners can now gather similar data at lower costs, higher frequency, and at higher fidelity. For the rest of the talk, I will go over the development of drones in recent history and how it will enable the next big data revolution, some example applications of drones in data gathering, and finally some challenges we are currently facing.
Unmanned aerial vehicles, commonly known as drones, were conceived almost as soon as airplanes were invented themselves. In less than 15 years after the first airplane was successfully flown by the Wright brothers in 1903, Hewitt and Sperry prototyped the first airplane that could be flown without a pilot. It used onboard gyros to stabilize itself, radios for remote control, and carried bombs to hit military targets, much like the drones today. The term drones became popular since the 1990’s when The United States Air Force developed and started using the Predator drones initially for intelligence gathering, and later for performing targeted strikes. The sensors they packed and the data they can gather were truly impressive. However, they were expensive, high maintenance, and required skilled pilots to operate them. A United States Air Force Reaper drone costs about $17M USD today with operating costs of thousands dollars per flight hour. Of course the collected data is classified and not available for civilian use.
https://outrunchange.com/2015/04/13/operating-costs-per-hour-for-usaf-planes/
In the past 5 years, as consumer drones have become more affordable, there has been a proliferation of drone applications. It is being used in almost every industry, from agriculture, construction, environment, to transportation, and real estate. Chris Anderson, CEO of 3D Robotics, recently said that you can apply drones to any industry, all of a sudden you get a new business. It is probably not an exaggeration to extend it even further and say that you can pick any word in the dictionary, add the word drone to it, and make a profit from the resulting business. Let’s take a look at a few examples of how different industries are using drones to gather data today.
Drone + Any Industry = New Business
Quote from Chris Anderson, CEO of 3D Robotics
Chris Anderson Interdrone 2015 Talk, September 2015
https://www.youtube.com/watch?v=1DqVl9P1FcY
Before the 1980’s, farmers typically applied the same amount of water and fertilizer to their crops based on their past experience. Because of the uniform application, some areas might be over fertilized, and some areas might not have enough. In 1985, researchers at the University of Minnesota developed the concept of dividing the farm into small grids and applied different amounts of lime to the crop. Precision agriculture is the practice of taking precise measurements of the soil, crop, and environment on indicators such as pH, moisture, crop maturity, etc., and based on these measurements, vary the amount of inputs such as water and fertilizer at the specific location where they are needed.
A drone carrying infrared and other sensors can quickly fly over and scan the entire field of crops to measure its growth and health. A farmer would have to otherwise drive to various parts of the field and visually inspect his or her crop, and might not notice problem areas that are easily visible through thermal imaging. A fleet of drones can completely automate this process by flying over the field periodically, imaging the entire field, and build a times series animation of the changes.
Not only do drones collect data, they can also carry fertilizers and pesticides onboard and spray them exactly where they are needed based on the prior analysis. The precise application of these chemicals leads to lower costs and better yield, and it’s great for the environment as well.
Drones are now used in the construction industry for monitoring and inspection, and in the not distant future, as a part of the construction process itself. Previously, inspectors had to climb high rises to physically measure every aspect of the building. Not only was this dangerous, it was also time consuming and expensive.
Nowadays, drones fitted with LIDAR or even standard cameras through 3D reconstruction techniques, can build precise 3D models of the buildings and do the measurements safely from the ground. As construction begins, project managers can use drones the monitor the status project and know early if certain areas are falling behind or not built to specification.
These fixed wing drones can also be used to protect the environment. They can fly for many tens of kilometers and loiter in the air for an hour or more.
It is estimated that over 30,000 elephants and over 1000 remaining rhinos are killed each year from poachers in Africa. Flown high up in the sky, authorities can use drones to track the poachers while avoiding detection. Further, they can also use infrared cameras to easily spot the poachers. Imagine uploading all of the images to the cloud, apply some deep learning magic, and you build a model that detects whether humans are tracking animals.
http://news.nationalgeographic.com/news/2014/08/140818-elephants-africa-poaching-cites-census/
One of the key reasons why drones are widely used and will eventually become ubiquitous is simply – cost. If you compare the cost and technical difficulty of operating a satellite, versus a plane, versus a drone, it’s quite apparent that drones are the most cost effective in many scenarios, and at the same time enable new data collection scenarios never before possible. A satellite costs hundreds of millions dollars to launch and maintain and is only available for large corporations. If a large farm wanted to image their farm, they would hire a pilot to fly over, and would costs thousands per day. At just a few thousand per drone, mom and pop farmers can afford the same level of data analytics as previously reserved for the corporate farms.
http://www.globalcomsatphone.com/hughesnet/satellite/costs.html
$50 to $400M USD to build and launch.
$3,500 USD per MHz of bandwidth per month, or a few million per year.
$180/hour plane only
https://www.quora.com/How-much-does-it-cost-to-own-a-small-plane
http://generalaviationnews.com/2011/06/08/flying-is-expensive-but-does-it-have-to-be/
Pilot $50-$100/hr
http://answers.google.com/answers/threadview/id/787273.html
https://answers.yahoo.com/question/index?qid=20110525131931AAQG5iJ
Just looking at enterprise drones alone, in five years there will be 800,000 units shipped globally. At $10,000 a unit, this is a $8 billion industry.
Our job as a drone manufacturer is then to build drones that not only affordable, but also safe and reliable, and autonomous. In addition, there must be ample regulatory support to allow widespread deployment of drones for commercial uses after they’ve been proven to be safe. If we can overcome the technical and regulatory hurdles, which I believe we can, then the industry is poised for an explosive growth.
Let’s go into a bit more technical details and see what’s really inside a drone. In the simplest terms, it’s nothing more than a cell phone with a bigger battery, four motors and propellers, and a larger camera mounted on a gimbal. A gimbal is itself a set of three motors that stabilize the camera from vibration and isolates the movement from the drone.
http://scottiestech.info/2015/06/21/lithium-polymer-vs-lithium-ion-batteries-whats-the-deal/
ARM 801 – Hover and Dobby
Same risks as cell phones – low margin business selling hardware
http://www.droneflyers.com/2015/12/metrics-and-indications-for-dot-drone-bust/
Many of the core components are shared with the cell phone. For example, the sensors its needs to maintain stable flight and navigate, the accelerometer, gyro, magnetometer, and GPS, are all there in a smart phone. The CPU is ARM based and runs Linux. Many drones use Wi-Fi and cellular networks to communicate with the ground station or servers on the Internet. Imaging is done through standard CMOS sensors and DSP’s developed for cell phones and small cameras. Power is provided by lithium ion battery packs, although a lot bigger.
http://scottiestech.info/2015/06/21/lithium-polymer-vs-lithium-ion-batteries-whats-the-deal/
ARM 801 – Hover and Dobby
Same risks as cell phones – low margin business selling hardware
http://www.droneflyers.com/2015/12/metrics-and-indications-for-dot-drone-bust/
Similar to Android today, there are complete open source hardware and software solutions for quickly building a drone and have it fly stably. First started as a project in ETH Zurich in 2009, the PX4 platform rapidly grew and was adopted by startups such as 3D Robotics. They demonstrated that it is possible to have commercial success based on this platform and nowadays, there are many dozens of startups in China alone trying to jump onto this bandwagon.
http://www.suasnews.com/2013/08/px4-and-3d-robotics-present-pixhawk-an-advanced-user-friendly-autopilot/
http://www.droneflyers.com/2016/03/3d-robotics-failure-to-launch-and-unicorn-dreams/
The same forces that drive the advance of cell phones year over year applies to the development of drone technology. Ultimately, we get the same user benefits such as lower cost, smaller form factor, and lower power. However, with higher volume and lower margins, any drone manufacturer that focuses exclusively on selling hardware alone will probably suffer the same fate as the manufacturers selling cheap Android phones today. There is a saying that the current crop of drone manufacturers will probably see a red ocean, meaning a blood bath, before they can reach the blue skies.
http://scottiestech.info/2015/06/21/lithium-polymer-vs-lithium-ion-batteries-whats-the-deal/
ARM 801 – Hover and Dobby
Same risks as cell phones – low margin business selling hardware
http://www.droneflyers.com/2015/12/metrics-and-indications-for-dot-drone-bust/
Let’s take a look at why the quadcopter design is so popular today and why it’s inherently safer than a fixed wing aircraft or a helicopter. First, it is mechanically simple. There are only four motors providing lift with no additional control surfaces or linkages between the motor and the control surface that could fail. In addition, quadcopters can perform vertical takeoff and landing with great wind resistance. By varying the amount of power provided by each motor, it can quickly change its orientation and direction of travel, making it extremely agile. By building smarter flight control algorithms, we can make the design even simpler and safer. Suppose that three of the four motors fail on the quadcopter. Can it still fly? In fact, Professor Raffaello D'Andrea recently showed that by spinning the drone around its vertical axis, one can still control the drone, albeit with reduced maneuverability. These qualities make it easy for anyone to control and fly quadcopter drones in almost any environment.
As we fly drones farther away and out of line of sight, and as more drones compete for airspace, we need to make them more autonomous. One such important feature is sense and avoid. It is the ability for drones to recognize obstacles in the environment, avoid collision, and navigate around them. Not only does it need to avoid static objects which may be difficult to see, such as power lines, it must also avoid moving objects such as other planes and birds. Similar to the autopilot feature in cars, although promising, it will still be a while before such technology is 100% reliable. The drones will need to do most of its collision avoidance processing onboard because it may not have access to the server. Further, it must react quickly and so the latency will need to be low. In the future, drones will be able to communicate with other drones nearby to actively avoid them.
Up until recently, the commercial drone business in the US has been hindered due to excessive FAA regulations and the slow process for approval. Every business that wanted to operate drones needed a Section 333 exemption, which took months to approve, and the operator must hold a manned aircraft pilot license. The release of Part 107 has completely changed what it means to operate a drone commercially. Set to take effect in this month, it relaxes some of the more onerous restrictions and enables faster adoption of drones for commercial use. However, many of the most compelling use cases for drones still won’t be permissible under the rule without a waiver, so in some respects the barn door hasn’t been completely thrown open. The restrictions on beyond visual line of sight operation in Part 107 still makes package delivery unfeasible, and the requirement for one operator for every drone means that the drones cannot be truly autonomous. While every country has different rules concerning commercial drone use, it is clear that proper regulatory support is needed for the continued growth of the industry.
Already, there are hundreds of companies involved in the drone ecosystem, from drone manufacturers, drone service companies, to companies whose sole mission is to shoot drones down from the sky where they’re not allowed to fly. Let me introduce a few companies whose focus is on big data.
DroneDeploy is a company started in 2013 that automates a lot of the flight path planning, data uploading to the cloud, and data processing. A user can simply draw the bounding box of an area they want the drone to cover, and DroneDeploy will figure out the optimal path that the drone should fly in order to cover the area with the captured imagery at the right resolution. They have algorithms for automatically stitching the images into a map and apply specific processing for example on thermal imagery.
How do you get that drone to actually make that growth map of your fields? It takes 50 to 70 clicks. Most of the time you must have a Windows laptop. You need to buy $8,000 pieces of software. It’s going to take 24 hours of processing time to actually make those images into the map that you’re looking for.
A user uses a mobile phone, turn on your drone, you press two buttons on your phone, the drone will launch it’ll capture the imagery, it’ll process that imagery and it’ll deliver you back that imagery, that map or that 3D model in a matter of minutes.”
The core feature offered by DroneDeploy at this point is tailored to its current customer focus of farmers, construction companies and mining companies: namely aerial imagery used for photogrammetry. So it can offer mapping and 3D modeling of large areas (of hundreds of acres) and provide volumetric estimates.
PrecisionHawk, started in 2011, is another company that offers similar services, but in addition they are trying to build an algorithm marketplace. Similar to an app store on the cloud, third party developers can write custom data analytics algorithms for specific industries and rent those modules to businesses who are in the industry. To me, it is clear that the future of the commercial drone industry is not just in the hardware, but also on the software and services. It is by building this platform and nourishing this ecosystem, is how drone companies will succeed and the industry will thrive.
Finally, let me discuss technical challenges for gathering, transmitting, and storing all of this data.
First, is the bandwidth required to transmit all of the data in real time. Currently, only navigational data are transmitted in real time, which take on the order of kilobits per second. Ideally, we would want to live stream all of the image and video data as well, which is on the order of megabits per second, and often in remote places. Imagine if every cell phone user were directly broadcasting from their phone today. That would quickly bring down current 4G networks.
Lidar – 100K points per second
Second, is the data transmission latency. Satellites are the only way to reach remote places, but they have a latency on the order of 1 to 2 seconds. Imagine if you’re trying to pilot a drone remotely in an inspection scenario, where you would have to get close to a building. If you don’t see the drone react until after two seconds you press a button, then it is very likely that the drone will run into a wall. For VR/AR scenarios, we are still very far away from minimum acceptable latencies for a good user experience.
http://www.evdoinfo.com/content/view/4818/64/
http://oculusrift-blog.com/john-carmacks-message-of-latency/682/
Third is the problem of having sufficient analytics algorithms to process the data for every industry. Even after the mom and pop farmers image their farm, they won’t have the expertise to stitch the images together or analyze the thermal images. Every industry might require their own custom processing, and this becomes a bottleneck in the ecosystem. Looking at the demand for computer vision engineers and their ever increasing salaries, it seems like we haven’t solved this problem yet. However, there are some techniques that are widely applicable. For example, image stitching and 3D reconstruction give a full picture of all of the captured imagery. The challenge then is to find similar techniques for other types of sensor data. The goal is to build a plug and play ecosystem of sensors, flight platforms, analytics platforms, to enable massive adoption for any industry.
Imagine a world where all of the drones are connected to the Internet and could be controlled from any PC or smart phone. The next obvious question is how do you let legitimate users to easily control and monitor the drones, while keeping the hackers out? While we still have a long way to go before there is SkyNet and a Terminator style apocalypse, it is not hard to imagine terrorist organizations wanting to take over the drones. Currently there is very little attention paid to this issue because most drones are controlled by a single operator through the remote control, we will soon need to solve this problem.
Let me talk a little about what we do here at EHang. EHang not just a drone company. Founded just 2 years ago, we have three main product lines on the market or in development.
On the consumer side, we have the GhostDrone which one of the first mass produced consumer drones controlled solely using a cellphone, without relying on remote controls. In addition, we have a VR goggle that gives the user a first person view of what it looks like up in the air. Our goal is to offer a completely immersive user experience. Flying the drone is as natural as a bird flying in the air.
Next we have the Falcon for the commercial drone industry. We call it an automated flight platform. Different industries can attach their custom sensors onto the drone and record the data. We have flight planning tools that where the user can easily specify exactly where they want to fly the drone, to minimize the risk of accidentally hitting something under manual controls.
Finally, we have the 184, the personal transportation drone. This is currently under development and the goal is to be able to provide point to point transportation for a single person.
I hope I have given you a glimpse of the future of the drone industry and the possibility it can bring. It is truly exciting to be in this space and I hope I have sparked your interest as well. Thank you.