DIGITAL MEDIA PROFESSIONALS INC. | Making the Image Intelligent






当社AI認識モデルの産業用ドローンへの応用 / Introduction of AI Recognition Model for Industrial Drones

当社AI認識モデルの産業用ドローンへの応用 / Introduction of AI Recognition Model for Industrial Drones


当社が開発したAI認識モデルは、ドローンが障害物にぶつからないようにする目的で使用しています。AI認識モデルは、ドローンの離陸時に前方の物体と物体までの距離を検出し、またドローンの飛行時は前方および前方下部の物体と物体までの距離を検出します。動画をご覧いただくとBounding Boxで物体検出「例: power_tower (送電塔)、person、turbine (風力発電用タービン)」し、Bounding Boxの下部にドローンからその物体までの距離を検出していることがおわかりになるかと思います。ドローンが離陸する際のカメラの角度による物体の見え方に関係なく、AI認識モデルは物体の種類を判別し距離を測ることが可能です。


距離推定の精度比較: 左写真 (DMP製)、右写真(他社製) /
Comparison of accuracy of distance estimation: Left photos(DMP), Right photos (Another)


We have developed an AI recognition model for industrial drones for distance estimation and object detection using a monocular camera.
This video shows an actual field test with a customer’s drone, in which NVIDIA Xavier was used as the operating environment.
The AI recognition model we developed is used to prevent the drone from hitting obstacles, which detects objects ahead of the drone and the distance to them when the drone takes off, and detects objects ahead of and below the drone and the distance to them when the drone is in flight. In the video, you can see that the Bounding Box detects objects, such as power_tower, person, and turbine, and the distance from the drone to the object at the bottom of the Bounding Box. Regardless of the angle of the camera at which the drone takes off, the AI recognition model is able to determine the type of object and measure the distance.

The feature of this AI recognition model is that it uses a monocular camera for distance estimation. The conventional orthodox method for distance estimation is to use a stereo camera. In this case, prior camera calibration is required and there are cost and size restraints. Our AI recognition model uses a monocular camera for distance estimation, which physically requires only one camera, thus saves space and reduces cost, and does not require any prior camera calibration. As long as you have a simple video, you can get pixel-perfect distance images and camera movement trajectories. In addition, as you can see in the image, it is possible to achieve highly accurate distance (depth) estimation compared to other AI networks that also use a monocular camera for distance estimation. Furthermore, combining it with the model that recognizes various objects such as cars, boats, pylons, and houses, you can estimate the distance to a specific object recognized.

This application to an industrial drone is just one example we can show, but distance estimation, together with SLAM technology, is a fundamental technology for realizing the robot’s eye in various applications such as self-driving cars, AGVs in factories and warehouses, UGVs, and cleaning robots. We are currently focusing on the development of technologies for autonomy and manpower saving in the field of robotics, and will contribute to the autonomy and automation of our customers’ robotic products with our advanced technologies such as SLAM technology and AI recognition models for distance estimation.