LeddarVision™

Combining Vayavision and LeddarTech Technologies

Bringing together the technologies from LeddarSense and VayaDrive software, the LeddarVision software platform combines AI and computer vision technologies as well as deep neural networks with computational efficiency to scale up the performance of AV sensors and hardware essential for planning the driving path.

Raw Data, Fused to Perfection

This state-of-the-art solution enables the detection of the various objects in the scene, including vehicles, pedestrians, bicycles, drivable road, obstacles, signs, lanes, lane lines, and more. LeddarVision also detects very small obstacles on the road with better detection rates and less false alarms than legacy “object fusion” solutions. Unclassified obstacles are also detected, providing an additional layer of safety to the vehicle.

• Comprehensive, open sensor fusion and perception platform
• Combines data from LiDAR, radar, and camera
• Provides LiDAR performance in Level 2 applications using only camera + radar
• Improves Level 3+ LiDAR performance and cost

• Most accurate environmental model
• Deep tech raw data fusion with upsampling
• Size, shape, and velocity information for every surrounding object
• Remarkably accurate and reliable detection and classification

Precise Object Detection

To detect objects, no single type of sensor is enough. Cameras don’t perceive depth, while distance sensors such as LiDARs and radars offer very low resolution. LeddarTech has just the solution: raw data fusion with upsampling.

Detection of small objects absent in training sets

Depth data assigned to every pixel in camera picture

Accurate shape definition of vehicles, humans, and any other object

Calibration, unrolling, and matching module receives raw sensor data before synchronizing and fusing it into a unified 3D model. Upsampling increases the effective resolution of the distance sensors, resulting in a dense RGBD model with each pixel containing both color and depth information. Localization and motion tracking help to determine the self-position and velocity.

Frame-based object detection and segmentation of the road includes obstacles, vehicles, pedestrians, bicycles, motorcycles, lane borders, available free space, and more. Detection by classification is performed with DNN-based algorithms that require training. In parallel, detection without classification is performed by another set of algorithms, thus enabling the detection of unexpected obstacles. Multiframe object tracking includes 3D modeling, motion tracking, and sizing of each object.

The resulting environmental model data is accessed via our software API and includes an occupancy grid and list of parameters for any tracked objects: localization, orientation, motion vector, and more.