WiMob2022
Two oral presentations at WiMob2022 in Thessaloniki, Greece
Others
International Conference · 2022
Object Recognition from 3D Point Cloud on Resource-Constrained Edge Device
Abstract
This paper presents the design and development of a lightweight, portable 3D spatial sensing device equipped with a compact LiDAR-type sensor. The device provides a 3D point cloud representation of the surrounding environment as captured by the sensor. Based on the acquired 3D point cloud, we propose a real-time object recognition method. The technical challenge is how to process the 3D point data in real-time while pursuing the best trade-off between the processing overhead and recognition accuracy. To answer the question, we leverage the Fisher Vector to extract spatio-temporal features of different objects enabling efficient classification of these objects using the Support Vector Machine approach. The experimental results show that the proposed method achieves mean Average Precision of 0.961. The processing speed on the device was 59.3 frames/second, indicating that object detection can be done in real-time on the device.
Research Note
Making three-dimensional spatial sensing portable removes the need for fixed installation and widens where it can be used. Small devices have limited computing resources, though, so processing point clouds on them in real time takes care. A point cloud carries a large number of points per frame, and deep learning applied directly to it does not keep up.
This work presents the design and development of a lightweight, portable 3D spatial sensing device equipped with a compact LiDAR-type sensor, which represents the surrounding environment as a point cloud. A real-time object recognition method is proposed on top of it, and the technical question is how to process the data in real time while striking the best trade-off between processing overhead and recognition accuracy.
The answer adopted here combines Fisher Vectors with a Support Vector Machine. A Fisher Vector encodes many local features into a single vector by taking gradients with respect to a probabilistic model, which reduces data with a varying number of points to a fixed-length representation. It yields discriminative features at far less computational cost than deep learning, which suits resource-constrained hardware, and here it extracts spatio-temporal features of the different objects.
The method achieved a mean Average Precision of 0.961, and the device processed 59.3 frames per second, showing that object detection can be performed on the device in real time.
Related Lab Activities
Related Projects
