The trend of electrification has become dominant, and intelligence is competing to bring more attractive intelligent functions to consumers, which will become the core competitiveness.
On May 9-10, 2023, at the 2023 Jietu Automotive Electronic Architecture and Intelligent Driving Forum, Wang Kun, co-founder and COO of Qingzhou Zhihang, stated that the trend of electrification has become dominant, and intelligence is competing to bring more attractive intelligent functions to consumers, becoming the core competitiveness.
Qingzhou Zhihang will adhere to standard configuration, standardization, popularization, and innovation, achieving a good high-speed NOA experience and achieving urban NOA point-to-point autonomous driving within a large ODD range in terms of user value for assisted driving.
Wang Kun | Co founder and COO of Qingzhou Zhihang
The following is a summary of the speech content:
Qingzhou Zhihang is an autonomous driving solution company established in 2019. Its core team comes from world-class autonomous driving companies and technology companies such as Waymo, Tesla, Nvidia, Facebook, Baidu, etc. The proportion of R&D personnel is as high as 80%, with nearly 80% of R&D personnel having master's and doctoral degrees. Employees with more than 5 years of relevant industry experience account for over 60%.
The company started in Silicon Valley and is mainly developing in China. Its technological foundation is forward-looking technology for autonomous autonomous driving, and based on this, it has launched products.
Reflection on the trend of intelligence in Qingzhou Intelligent Navigation
The following picture shows the "dual engine strategy" of Qingzhou, and on the left is the power engine. The relevant product solutions that will implement forward-looking technology at the L4 level, including unmanned minibuses in 10 cities across the country; On the right is the innovation engine, which applies forward-looking technology to mass production solutions. Currently, it is also working with some OEMs to scale up production and create L2+level solutions. Below is the technology base, namely the autonomous driving super factory, which includes a light boat matrix of data feedback and data closed-loop, the progress of iterative technology, and the generation of product solutions.
Source: Qingzhou Zhihang
The trend of electrification has become dominant, and intelligent competition has emerged. We have also proposed the small four modernizations of light boats:
We hope to quickly deploy the basic functions of intelligent driving to various vehicles and make it standard for the first entry-level experience.
The standardization of the second mid-range experience is to push the functions of high-speed NOA related products towards a more standardized experience, and users know what expectations they should have after the functions are installed in the car.
The popularization of the third high-end experience not only applies more advanced functions to high-end models through technological development and progress, but also allows mid to low-end models to experience urban NOA products.
The formatting of the fourth ultimate experience reduces the dimensionality of innovative technologies, including those from L4 level to L2 level, bringing a product experience to the public.
Upgrade the capability of intelligent assisted driving solutions
The following figure is a product function diagram. We have circled the scenarios covered by different levels of intelligent driving functions. Cruise control and track keeping assistance on highways are common L2 functions, while high-speed NOA belongs to L2+and will add functions such as automatic lane changing, high-speed gate diversion, merging, and intelligent avoidance. Cities are more complex, including traffic light entrances on urban surface roads, U-turn turns, narrow lane traffic, and parking scenarios.
Source: Qingzhou Zhihang
We believe that the value of assisted driving lies in being better and more loving to use. Better refers to meeting the basic high-speed NOA experience and achieving point-to-point driving in some urban areas, with lower driver takeover rates and providing a more reassuring and trustworthy butler style experience. Preferably, it refers to achieving point-to-point autonomous driving of urban NOA within a larger ODD range, achieving more simple autonomous driving, making driving as easy as taking a taxi, making drivers dependent on assisted driving, and enjoying the driving process more easily.
At the recent Shanghai Auto Show, the mainstream intelligent driving solutions for mid range models were roughly divided into two categories: one based on LiDAR, and the other focused on vision. The general configuration is based on computing chips, mainly Orin and Horizon.
Source: Ruisiqi Consulting
Intelligent assisted driving solution for Qingzhou Intelligent Navigation
Based on market demand, Qingzhou has proposed two solutions. One is a 1L11V5R solution that includes high-speed and urban NOA, supporting single/dual journey 5 chips. To achieve high cost-effectiveness, it is necessary to optimize single and dual J5, and sensors are also the direction of optimization. The second set has a higher cost-effectiveness and is based on a 6V1R solution with high-speed NOA+L2 functionality. This solution is a more traditional front view, rear view, and four surround view camera, currently mainly implemented on a single J5 chip.
The following is a simple configuration diagram. Qingzhou has become one of the leading providers of urban autonomous driving solutions based on domestically produced high-performance "Journey 5" chips. In collaboration with Horizon, it is equipped with the first domestically produced 100 TOPS large computing power chip to achieve front-end mass production, creating a mass-produced intelligent driving solution that is functional, easy to use, and easy to use.
Source: Qingzhou Zhihang
At the same time, the full stack algorithm of Qingzhou is highly compatible with the NVIDIA Orin DRIVE platform and received official recognition for the cooperation between the two parties at the 2022 NVIDIA GTC conference. The center of the figure below is a display of Qingzhou's dual Orin solution based on NVIDIA.
Source: Qingzhou Zhihang
Leading engineering capabilities accelerate the landing of light boat products
Why can Qingzhou achieve a product with a good user experience? The essence lies in the characteristics of the solution: based on data; Become perceptive; Proficient in Pnc.
Firstly, regarding perception, we propose the concept of hyper fusion, which involves multi-sensor temporal interpolation fusion, including inputs from Lidar, Radar, and Camera. Temporalization can integrate past sensor inputs into the network, resulting in more accurate outputs. The entire large model is called OmniNet, which is based on the feature space of BEV, combined with multi-sensor fusion. Finally, through the shared backbone network Backbone, multi-sensor features are fused, combined with temporal characteristics, to output a model of multiple perception tasks.
In the following figure, BEV results are shown on the left, and the car identified by aerial image can be seen. This is a LiDAR recognition effect that can achieve visual perception results from 150-200 meters. The intermediate results can be directly output, including 2D lane lines, panoramic views, depth estimation effects, and segmentation effects. Finally, multi task complementarity can be obtained, resulting in the final perception result, which can be used for decision-making and planning by the control team.
Source: Qingzhou Zhihang
The left side of the figure below shows that after removing Lidar, the network design can still support output. After the input of six cameras under the whole aerial view, the vehicle's visual output, including the output of expressway and expressway, and the output of different lane lines, can be output on the network.
Source: Qingzhou Zhihang
The output of urban roads, including depth, detection and segmentation of visual obstacles, and geometry of lane lines, can be used in vision to solve crisis obstacle detection; In the tunnel, its perceptual output results are equally stable.
OmniNet provides rich and accurate environmental perception results, which can effectively improve the accuracy and accuracy of perception; At the same time, following a design that better utilizes the computing power of on-board chips, optimizing task performance through input and output can save 1/4-1/5 of the computing power compared to single tasks. The perception model can also be configured according to different sensors. As long as the final configuration and data collection process are newly determined, it can adapt to the entire vehicle model, achieve high-speed adaptation and low switching costs, and achieve the original purpose of high cost performance.
In addition to perception, Qingzhou has also launched an industry recognized superior spatiotemporal joint algorithm. The common or commonly used method in the industry is the spatiotemporal separation algorithm, which is the process of optimizing paths and speeds separately. For example, red represents a self driving vehicle. To overtake or change lanes, the path will be planned first, and then the speed at which to pass at different times will be planned under the path. The disadvantage is that if the path planning reveals that the vehicle cannot pass, the possible action to take is to brake first and then avoid, or if the lane change fails.
Source: Qingzhou Zhihang
The advantage of spatiotemporal joint optimization is that the vehicle prediction model is also included, not only considering the path of the vehicle itself, but also taking into account the opponent's action path. After comprehensive consideration, acceleration and overtaking can be done to solve more complex scenarios.
Of course, good planning also relies on good prediction models. Our self-developed prediction model, ProphNet, won the championship in the 2021 World Cup Armovers International Prediction Challenge in the field of autonomous driving and finished third in 2022. The entire prediction algorithm has achieved a prediction length of 10 seconds, can support hundreds of targets, and the inference time is less than 20 milliseconds. The J5 platform has also been optimized, with an accuracy of over 90% for Chinese intersection entry and exit models.
In addition to algorithms, perception, control, and prediction, data efficiency is also crucial as it determines the efficiency of algorithm iterations. Data platform, annotation platform, training platform, simulation platform, all algorithms are quickly circulated on the platform to iterate algorithm effects. We have also built a closed-loop tool chain that quickly solves problems from real vehicle testing, simulation testing, data processing, to actual simulation scenarios and virtual simulation scenarios, including training on annotated data.
Source: Qingzhou Zhihang
We can quickly deploy and train a new model within a week, while mining more similar data on the platform to quickly improve the problem-solving ability of corner cases.
(The above content is from Wang Kun, co-founder and COO of Qingzhou Zhihang, who delivered a keynote speech on "How to Create Better and Consumer Favorable Intelligent Driving Solutions" at the 2023 Jietu Automotive Electronic Architecture and Intelligent Driving Forum from May 9th to 10th, 2023.)
AMS2024 Exhibition Guide | Comprehensive Exhibition Guide, Don't Miss the Exciting Events Online and Offline
Notice on Holding the Rui'an Promotion Conference for the 2025 China (Rui'an) International Automobile and Motorcycle Parts Exhibition
On September 5th, we invite you to join us at the Wenzhou Auto Parts Exhibition on a journey to trace the origin of the Auto Parts City, as per the invitation from the purchaser!
Hot Booking | AAPEX 2024- Professional Exhibition Channel for Entering the North American Auto Parts Market
The wind is just right, Qianchuan Hui! Looking forward to working with you at the 2024 Wenzhou Auto Parts Exhibition and composing a new chapter!
Live up to Shaohua | Wenzhou Auto Parts Exhibition, these wonderful moments are worth remembering!
Free support line!
Email Support!
Working Days/Hours!