Mogo Auto launches BalanceHRNet vision-based autonomous driving estimation algorithm

Shanghai (Gasgoo)- For an autonomous driving system to handle sudden events and respond appropriately to pedestrians, it must understand common gestures and body languages. This understanding increases the safety and efficiency of autonomous driving.

Recently, Mogo Auto, a Chinese autonomous driving tech developer, introduced the BalanceHRNet human body posture estimation model, which the company says can accurately identify pedestrian intentions. This research has been included in the international top journal Neural Networks.

Neural Networks is one of the world's three authoritative academic journals in the field of neural networks. It is also an SCI Zone 1 journal, covering behavior studies, brain modeling, learning algorithms, mathematical and computational analysis, and the engineering and technical applications using neural network concepts and techniques.

Identification effects by BalanceHRNet model; photo credit: Mogo Auto

At present, the mainstream method for human posture estimation is to use neural networks to identify human keypoints. In recent years, researchers in this field have proposed various deep learning models to improve the accuracy of human posture estimation, including the mainstream HigherHRNet model. Although this model has made great progress, the identification accuracy can greatly decrease in complex environments or crowded scenes due to the mutual obstruction between different individuals.

To address the aforesaid issue, Mogo Auto proposed the BalanceHRNet model. This model draws on the multi-branch structure and fusion method of HigherHRNet, overcoming the inability of HigherHRNet to obtain a large field of view, thus improving accuracy with smaller computational complexity.

According to the company, BalanceHRNet has three major advantages: it possesses a larger field of view, extracting richer semantic information with higher accuracy; it proposes a Balanced High-Resolution Module (BHRM) to obtain multi-scale features of objects; it learns the importance of different branches, letting the model decide the significance of each branch itself.

According to Mogo Auto’s introduction, the researchers used the CrowdPose dataset as the test dataset, and models such as HigherHRNet, AlphaPose, and OpenPose as comparison models. The data showed that BalanceHRNet achieved an average accuracy rate of 63.0%, an increase of 3.1% compared to the model HigherHRNet. Although the per-frame perception accuracy only increased by 3.1%, this number has a big impact on the overall improvement of the safety of autonomous driving.

Researchers also demonstrated the effectiveness of the BalanceHRNet network through the COCO (2017) keypoint detection dataset. The BalanceHRNet model improved the average accuracy rate by 1.6%.

Currently, BalanceHRNet has been used by Mogo Auto in high-precision maps, improving the success rate of perception and map accuracy. The company said it will continue to promote technological breakthroughs in the fields of digital transportation and autonomous driving.