AI Perception Algorithm

Overview
Hirain is committed on researching and landing of Artificial Intelligence (AI) related technologies, providing solutions for perception and localization problems of self-driving vehicles in complex real  scenarios. Based on years of technical reserves in the field of automotive electronics, engineers in Hirain have developed a series of AI perception algorithms, which applied in varieties of autonomous driving scenes through in-depth analysis of multiple sensors such as Lidar, camera, millimeter-wave Radar, differential GNSS, high-precision maps and inertial measurement unit.
 

Construction of AI Perception Algorithm

In order to ensure the reliable implementing of AI perception algorithms, Hirain explores and forms a complete set of algorithm development and iteration process. A complete tool set, with according to methodologies, has been developed in this process, providing solutions from data acquisition, data labeling, data cleaning, model training, parameter optimization, pruning, quantification, compression, and embedded platform deployment.

Lidar based Point Cloud Object Detection and Semantic Segmentation
In a Self-driving scene, Lidars, with advantages like high resolution, high anti-jamming ability, abundance of information obtained and irrelevancy to the lighting environment, can complete the acquisition of three-dimensional spatial data effectively. However, Lidars still have shortcomings such as disorder and sparse of data collected.
Our AI Perception Algorithm, by extending the ability of deep neural network in data expression, combined with network design and parameter optimization, has accomplished the goal of traffic objects detection and road environment identification. Furthermore, this algorithm has been deployed on an embedded platform, with real-time, high detection accuracy and robustness.
The algorithm tool set has already been successfully applied on several off-road Self-driving Vehicle demonstration projects. Meanwhile, it has also been fully applied and verified in our Intelligent Port project.

Image based object detection and Semantic Segmentation
Teams in Hirain have carried out in-depth research and development on pedestrians, vehicles, ground signs and traffic lights for the traffic conditions of different autonomous driving scenes. Since visual sensors always return more features relevant to color, on the basis of cutting-edge technology, the performance of detection and semantic segmentation in various traffic scenes can be further improved.
By selecting the appropriate deep learning model and on-board embedded computing equipment from the point of view of mass production, the training and deployment of the model is realized by collecting, analyzing and labeling a large amount of data. Meanwhile, the test result analysis, model adjustment and performance improvement are carried out from the actual application scene. As a result, the tool set is finally applied in several Self-driving Vehicle demonstration projects.
 

High-precision Localization System
The Localization System developed by Hirain provides accurate and reliable localization information for autonomous driving systems by combining inputs from multiple sensors (visual sensors, Inertial Measurement Units, odometers), high resolution maps, high accuracy GPS and road signs (including traffic signs, lane lines, and road boundaries) identified by AI-perception algorithms.
 

Summary
The AI perception algorithm developed by Hirain provides significant support for the perception and localization of self-driving vehicles. By utilizing the features of different sensors, we designed and trained different models targeting for stable operation needs in different self-driving scenes as well as different weather. At the same time, our automotive grade and low power consumption solutions ensure the reliability and competitiveness of the AI perception algorithm.