Wiring brake controllers for a car to tow a trailer can be tricky, and they can get damaged.
Instead, let's make trailers smarter using small cameras and computers. We can use computer vision techniques to detect the actions of a tow vehicle's lights based on the brightness, location, and changes in these values over time.
Then, we can use that data to control trailer lights and brakes. This eliminate the problem of altering the wiring of a tow vehicle. In the future, we can add more computer vision capabilities.
- Vehicle light configurations vary
- Weather conditions vary
- Vehicle pose varies
- Other lights will be visible beyond the envelope of the tow vehicle
- Being a safety critical application, a safety testing framework must be constructed
- Failsafe behavior must achieve a minimum risk condition
- YOLOv8
- Python
- Norfair
- Boolean logic
Research on intelligent vehicle lamp signal recognition in traffic scene, Shi, Qi, Liu, Yang
https://link.springer.com/article/10.1007/s42452-022-05211-9
This approach uses YOLOv3 to detect vehicles and then uses hue-saturation-value data to improve detection. The data is fed into a DNN for processing and detection tasks.
Autonomous Tracking of Vehicle Rear Lights and Detection of Brakes and Turn Signals - Almagambetov, Casares 2012
https://www.researchgate.net/publication/235676076_Autonomous_Tracking_of_Vehicle_Rear_Lights_and_Detection_of_Brakes_and_Turn_Signals
Use of Kalman Filter and Codebook to achieve very good signal detection performance.
Vision-based Driver Assistance Systems: Survey, Taxonomy and Advances: Horgan, Hughes, McDonald, Yogamani 2021
https://arxiv.org/pdf/2104.12583
Introduces taxonomy for ADAS, includes review of historical approaches and benchmarks/limitations.
Takeaways:
30hz for low speed maneuvering cameras
15hz at night for longer exposure time.
Early methods: Hough Transform, Canny Edge Detection.
CNNs displaying better performance in detection/segmentation 2020 onward.
Pixel segmentation: Labels each pixel into category.
Long et al 2015 showed FCN efficacy for semantic segmentation, allowing ADAS to detect fine details
DNN & RL great for making decisions from large volumes of data
Region-based CNN great for object detection & collision avoidance
Key problem is sensor fusion
Must integrate camera, LiDAR, ultrasonic, radar; cameras suffer in poor visibility / illumination
Kalman filter principal method for fusing inputs
Real-time tracking key issue to to compute constraints
Single Shot multibox Detector SSD