Regression Based Traffic Sign Detection using Transfer Learning For Self Driving Cars
10th Annual COE Graduate Poster Presentation Competition
Enoch Sarku (PhD)
Balakrishna Gokaraju, Ali Kharimodini, Sun Yi
- Humans find it relatively easy to recognize traffic signs using our eyes and brains. However, machines see traffic signs as a bunch of as 0’s and 1’s. Traffic sign detection can help autonomous cars increase safety requirements set out by the Federal Automated Vehicles Policy (FAVP). Traffic Sign Detection accuracy could be improved using Deep Learning Technique
- However Deep Learning Techniques require huge amounts of data resulting in high cost of data i.e. more labor, time and resources required and at least 1000 images [1].
- A traffic sign model must operate a pipeline that minimizes computing and business costs
- Proprietary datasets released for public use are not intentional towards targeting traffic signs e.g. Waymo dataset, Uber & Lyft
- Traffic Sign Detection is a complex classification problem [3]
- Traffic signs are critical for vehicle and pedestrian safety
- Traffic sign models must be as accurate as possible
Figure 1: Yolov5 Arcjitecture pipeline
- Data collected manually using autonomous vehicle camera, video to image convertor used to extract images
- 16 classes of traffic signs including stop and speed limit signs, were selected with some degree of imbalance
- You Only Look Once (Yolov5) model was used to train traffic sign objects
- Yolov5 employs a Regression based approach layered with a pretrained model such as Residual Net.
Figure 5: Shows performance of model after training on 1000 images
Figure 6: Collection process for the lane detection model
The model was deployed using the robot Operating system pipeline. The model performed well in scenes that were similar to the training data and struggled fairly on completely variable data.
- [1] Bochkovskiy A., Wang C. YOLOv4: Optimal Speed and Accuracy of Object Detection. arXiv. 20202004.10934v1
- [2] J. Hochstetler, R. Padidela, Q. Chen, Q. Yang, and S. Fu, “Embedded deep learning for vehicular edge computing,” Proc. - 2018 3rd ACM/IEEE Symp. Edge Comput. SEC 2018, pp. 341–343, 2018, doi: 10.1109/SEC.2018.00038.
- [3]F. Fahimi, Autonomous Robots. 2009.
- To detect traffic signs using bounding boxes
- To account for varying distance, partial occlusion and size properties associated with traffic signs.
Limitations & Future Work
The previous work did not account for major problems associated with traffic sign detection and classification models. Scenarios such as Occlusion and Partial Occlusion, multiple signs in an image, shear and distortions and many more. Future work will focus on addressing these issues by incorporating them into the training model.
The model detected and classified 16 traffic signs include stop sign, crosswalk and speed limit sign among others. For most signs the model was between 75% to 89% confident of its classification. Especially for stop signs, a false detection could lead to devastating consequences for self driving vehicles.