Creation of algorithms for 3D spatial-temporal data modeling

Object recognition using 3D data
Object recognition using LiDAR intensity
Stereo self-rectification in natural scenes

Motivation: Realizing human-like perception mechanism composed of “object separation” and “classification using silhouette shape”

Aimed function: Detection, recognition and object tracking

Application: Autonomous driving, ADAS, surveillance camera and sports analytics

Watch video on Youtube : https://youtu.be/niWcTAag8uY

Separate all objects FIRST (inspired from human perception mechanism). All objects are separated and then their silhouette is extracted
Handle occlusions: deal with “True/Fake contour” distinctively, this system achieves good recognition even if an occlusion exists
By taking into account the continuity between frames, this system reduces FP/FN
Simple, compact silhouette descriptor of fixed size, easy to classify easy to handle, to transmit…1-2 layers NN, does not require GPU (not like CNN)
Training from few samples due to silhouette uniqueness
Accurate shape expression, even the hand in the pocket could be characterized. Silhouette technology allows easy gesture and human pose recognition

SIlhouette is a robust descriptor allowing high accuracy and real-time processing

(*) images are not used, LiDAR model : Velodyne – 16 planes – H=1.8m

(*) images are not used, LiDAR model : Velodyne – 16 planes – H=1.8m

PROCESSING < 100ms/Lidar scan CPU: Cortex-A17, 4 cores, 1.80 GHz, RAM: 2GB, GPU: Mali-T760MP4 (no need)
Technology could be applied for classification of various kind of objects such as cars, cycles, poles, sitting persons…