Zhijia Technology's paper DualBEV was selected into the top computer vision conference ECCV

Recently, the paper "DualBEV: Unifying Dual Veiw Transformation with Probabilistic Correspondences" by the Zhijia Technology team was selected into the European Conference on Computer Vision (ECCV). ECCV is the most influential and authoritative international conference in the field of computer vision. One, it is as famous as the International Conference on Computer Vision (ICCV) and the Conference on Computer Vision and Pattern Recognition (CVPR), and is known as one of the "three top conferences" in the field of computer vision. ECCV is held every two years and brings together the world's top researchers and experts to showcase and discuss the most cutting-edge research results and technological innovations.

DualBEV: The best BEV perspective conversion method on the list

Bird's-Eye-View (BEV) perception is the cornerstone of current autonomous driving perception and even end-to-end technology, among which perspective transformation (View Transformation) is the core of BEV perception Module, responsible for converting image features (2D) to BEV space (3D). Current mainstream solutions often fall into a dilemma between 3D-to-2D or 2D-to-3D perspective conversion solutions.

3D-to-2D solutions generally rely on Transformer, which while achieving good performance also brings a lot of computing overhead. Although 2D-to-3D is fast in calculation, it is easy to lose long-distance information that is important in truck scenes.

In response to these problems, DualBEV started from the concept of Monte Carlo, and through thinking about the essence of View Transformation, summarized the process of View Transformation into constructing samples and calculating weights, and then proposed a universal feature transformation algorithm (Unified Feature Transformation ). The algorithm first constructs samples from two directions, evaluates the samples constructed in two different directions uniformly through the multiplication of three probability measurements, and then accelerates the conversion process through pre-calculation, and fuses them to obtain the final BEV features.

Universal feature conversion algorithm

DualBEV pioneered the unification of 3D-to-2D and 2D-to-3D perspective conversion solutions into an overall framework, taking full advantage of the advantages of each perspective. Achieved SOTA results with 63.4% NDS on the nuScenes Detection Leaderboard purely visual solution. Since the perspective conversion module uses pre-computation technology, its calculation time is only 1/40 of the Transformer solution, making it the best perspective conversion method on the list.

Through efficient perspective conversion and multi-view information fusion, the BEV features built by DualBEV provide accurate scene representation and fast calculation solutions, providing a solid foundation for the development of end-to-end systems. Zhijia Technology is actively promoting the deep integration of DualBEV in the end-to-end autonomous driving system, giving full play to its advantages and further improving the overall performance of the autonomous driving system.

Technology-enabled products, open source cooperation and win-win results

DualBEV’s further optimized related technologies have been applied to Zhijia Pilot 2.0, a pre-installed mass-produced heavy truck autonomous driving system independently developed by Zhijia Technology. The system has opened up the pilot's autonomous driving function, which has the advantages of extreme safety, comfort and labor saving, energy saving and environmental protection. It can realize advanced functions such as autonomous overtaking, parking at the side, merging prediction and partial construction avoidance, and supports "dual driving" in typical express delivery scenarios. "Change to single driving" to significantly reduce driving fatigue and achieve energy conservation and emission reduction through fuel savings of up to 10%.

The smart heavy-duty truck K7+ currently equipped with the Smart Plus Navigation 2.0 system has been put into actual operation by leading logistics companies such as China Post, ZTO Express, and Aneng Logistics, covering core economic zones such as Beijing-Tianjin-Hebei, the Yangtze River Delta, and the Pearl River Delta. Continue to empower the digital and intelligent transformation of the trunk logistics industry.

This paper was successfully selected into ECCV and some of the research results were open sourced. It not only demonstrates the scientific research and innovation capabilities of Zhijia Technology in the field of autonomous driving, but will also help improve industry standardization and interoperability. Zhijia Technology hopes to focus on overcoming key technical problems through innovation, cooperation and sharing, effectively promote the iteration and breakthrough of autonomous driving technology, and ultimately achieve high-quality development of "intelligent, safe and environmentally friendly" logistics and transportation industry.

The above is the detailed content of Zhijia Technology's paper DualBEV was selected into the top computer vision conference ECCV. For more information, please follow other related articles on the PHP Chinese website!