Training neural networks for automotive computer vision systems considering types of false estimations

Introduction (problem statement and relevance). This article deals with the problem of training artificial neural networks intended to analyze images of the surrounding space in automotive computer vision systems. The conventional training approach implies using loss functions that only improve the overall identification quality making no distinction between types of possible false predictions. However, traffic safety risks associated with different types of prediction errors are unequal being higher for false positive estimations.

The purpose of this work is to propose improved loss functions, which include penalties for false positive predictions, and to study how using these functions affects the behavior of a convolutional neural network when estimating the drivable space.

Methodology and research methods. The proposed loss functions are based on the Sørensen-Dice coefficient differing from each other in the approaches to penalizing false positive errors. The performance of the trained neural networks is evaluated using three metrics, namely, the Jaccard coefficient, False Positive Rate and False Negative Rate. The proposed solutions are compared with the conventional one by calculating the ratios of their respective metrics.

Scientific novelty and results. The improved loss functions have been proposed to train computer vision algorithms featuring penalties for false positive estimations. The experimental study of the trained neural networks using a test dataset has shown that the improved loss functions allow reducing the False Positive Rate by 21%.

The practical significance of this work is constituted by the proposed method of training neural networks that allows to increase the safety of automated driving through an improved accuracy of analyzing the surrounding space using computer vision systems.

1. Winner H., Hakuli S., Lotz F., Singer C. Handbook of Driver Assistance Systems. Basic Information, Components and Systems for Active Safety and Comfort. Springer International Publishing Switzerland, 2016.

2. Pearl J. Probabilistic reasoning in intelligent systems: Networks of plausible inference. San Francisco, Morgan Kaufmann Publishers inc., 1988, p. 1.

3. Duda R., Hart P. Pattern Clas sification and Scene Analysis. Hoboken, John Wiley and Sons, 1973.

4. Goodfellow I., Bengio Y., Courville A. Deep Learning. Massachusetts, MIT Press, 2016, p. 13.

5. Cireşan D.C., Giusti A., Gambardella L.M., Schmidhuber Jü. Deep neural networks segment neuronal membranes in electron microscopy images. Proceedings of the 25th International Conference on Neural Information Processing Systems, 2012, pp. 2843–2851.

6. Krizhevsky A., Sutskever I., Hinton G.E. Imagenet classification with deep convolutional neural networks. Adv. Neural Inform Process Syst., 2012, pp. 84–90.

7. Cityscapes Dataset Benchmark Suite. Available at: www.cityscapes-dataset.com/benchmarks/ (accessed 26 May 2021).

8. KITTI Vision Benchmark Suite. Available at: www.cvlibs.net/datasets/kitti/eval_road.php (accessed 26 May 2021).

9. Rumelhart D., Hinton G., Williams R. Learning representations by back-propagating errors. Nature, 1986, pp. 533–536

10. Kingma D.P., Adam Ba J. A Method for Stochastic Optimization. Available at: https://arxiv.org/abs/1412.6980 (accessed 26 May 2021).

11. Ronneberger O., Fischer P., Brox T. U-Net: Convolutional Networks for Biomedical Image Segmentation. Available at: https://arxiv.org/abs/1505.04597 (accessed 26 May 2021).

12. Hughes C., Chandra S., Sistu G., Horgan J., Deegan B., Chennupati S., Yogamani S. Drivespace: Towards context-aware drivable area detection. Electronic Imaging, 2019, pp. 423–428.

13. Pizzati F., Garcia F. Enhanced free space detection in multiple lanes based on single CNN with scene identification. Proceedings of 2019 IEEE Intelligent Vehicles Symposium (IV), 2019, pp. 2536–2541.

14. Baheti B., Innani S., Gajre S., Talbar S. Eff-Unet: A novel architecture for semantic segmentation in unstructured environment. Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 1473–1481.

15. Milletari F., Navab N., Ahmadi S. V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation. 2016 Fourth International Conference on 3D Vision (3DV), 2016, pp. 565–571.

16. Paszke A., Chaurasia A., Kim S., Culurciello E. ENet: A Deep Neural Network Architecture for Real-Time Semantic Segmentation. Available at: https://arxiv.org/abs/1606.02147 (accessed 26 May 2021).

17. He K., Zhang X., Ren S., Sun J. Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification. Proceedings of IEEE International Conference on Computer Vision, 2015, pp. 1026–1034.

18. Cordts M., Omran M., Ramos S., Rehfeld T., Enzweiler M., Benenson R. The cityscapes dataset for semantic urban scene understanding. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 3213–3223.