Recently, through large-scale data accumulation at the vehicle end and road end, Juefei Technology has proposed innovation and optimization based on the MapTR (Map TRansformer) method: ① Optimize the expression of lane information and simplify the model structure; ② In the MapTR Based on the prior information of the map, the accuracy and recall rate of the model output map elements are effectively improved; ③The modeling of elements such as lane centerlines and road topology is added to systematically improve the efficiency of real-time map building for single vehicles, which is easy for automatic driving regulation and control use.
In 2022, the Vision Lab of Huazhong University of Science and Technology cooperated with Horizon and jointly proposed MapTR, an online real-time construction method for high-precision vector maps. It performs structural modeling on the map elements, and represents the map elements as a set of equivalently arranged point sets, which eliminates the ambiguity in the representation and reduces the learning difficulty of the model. At the same time, hierarchical query vectors are used to encode map structured information, which can flexibly encode point information and instance-level information.
Experiments show that MapTR has achieved good mapping quality and real-time running speed on the nuScenes dataset, and can maintain stable mapping performance in diverse and complex driving scenes. MapTR shows the potential and prospect of online mapping scheme, which has high application value.
The optimization direction of Juefei Technology: more suitable for downstream tasks of autonomous driving
Real-time mapping capability is a prerequisite for realizing real-time environment perception for autonomous driving. At present, the map vectorization method has become a hot direction for real-time mapping. It uses an ordered point set to represent each map element, and directly returns the point set of each map element, achieving more accurate results and faster operation speed. .
However, in practical applications, such methods also need to be improved. For example, when detecting multiple road information, the model structure is redundant, resulting in a decrease in computational efficiency; for occlusions and distant targets, effective features cannot be extracted. At the same time, this method cannot directly output the topology structure of the lane level, which makes it difficult for downstream regulation modules to use.
In response to the above problems, Juefei Technology has optimized the MapTR method for practicality through the accumulation of a large amount of vehicle-side and road-side data. Through these optimizations, it can more accurately capture map details and structures, and further enrich the output of map elements from the level of map construction. Making it more suitable for various downstream tasks of autonomous driving.
Juefei Technology's ability to perceive vector maps in real time
Video: Juefei Technology's Real-time Perception Vector Map Capability Demonstration
Practice of Juefei Technology
1. A more complete way to express lane attributes
On the basis of MapTR, Juefei Technology has increased the expression and output of the centerline of the lane, which can reflect the connection relationship of the lane more clearly. At the same time, Juefei has also increased the ability to learn the direction of the lane, and distinguishes the direction information of the centerline of the lane from the direction of the opposite lane. Not only that, but the model can also return lane width information and lane boundary line attribute information. This way of expression is more in line with the real needs of the downstream regulation and control module, and greatly improves the performance of the self-vehicle for map understanding.
2. Add map prior information
Through the accumulation of a large amount of road test data, Juefei Technology has added map prior information to MapTR, especially in the scene of "intersection" or occlusion relationship, this method can output the topological connection relationship more stably, which greatly improves the Accuracy and recall rate, thereby improving the safety of the self-vehicle when passing through complex intersections and other scenes.
3. Optimization of road topology expression
Lane objects detected by MapTR are usually interrupted by topological bifurcation points. Juefei Technology’s solution can re-attach the lanes interrupted by bifurcation points, without the need for subsequent manual processing of lane attachment. Although this method will lead to partial duplication of detected lanes, it can make each lane more independent and the lane shape more continuous and smooth, especially in scenarios where lanes often fork, it can provide a clearer supervision signal for the model, Improve the stability of the overall model used in actual scenes.
At present, Juefei Technology is adding multiple improvements based on MapTR to multi-task joint training, combined with 3D obstacle detection, road segmentation and other capabilities, and continuously iterates the large perception model to endow vehicles with better automatic driving capabilities in unfamiliar scenes . Juefei Technology hopes that this method can provide a basis for further innovation in real-time mapping, and ultimately improve the development of autonomous driving technology in a safer and more reliable direction.
Link to the original paper: https://arxiv.org/abs/2208.14437
Code link: https://github.com/hustvl/MapTR