Automatic Vehicle Ego Body Extraction for Reducing False Detections in Automated Driving Applications
Ciarán Hogan, Ganesh Sistu · Communications in computer and information science · 2023
Abstract Fisheye cameras are extensively employed in autonomous vehicles due to their wider field of view, which produces a complete 360-degree image of the vehicle with a minimum number of sensors. The drawback of having a broader field of view is that it may include undesirable portions of the vehicle’s ego body in its perspective. Due to objects’ reflections on the car body, this may produce false positives in perception systems. Processing ego vehicle pixels also uses up unnecessary computing power. Unexpectedly, there is no literature on this relevant practical problem. To our knowledge, this is the first attempt to discuss the significance of autonomous ego body extraction for automobile applications that are crucial for safety. We also proposed a simple deep learning model for identifying the vehicle’s ego-body. This model would enable us to eliminate any pointless processing of the car’s bodywork, eliminate the potential for pedestrians or other objects to be mistakenly detected in the car’s ego-body reflection, and finally, check to see if the camera is mounted incorrectly. The proposed network is a U-Net model with a Res-Net50 encoder pre-trained on ImageNet and trained for binary semantic segmentation on vehicle ego-body data. Our training data is an internal Valeo dataset with 10K samples collected by three separate car lines across Europe. This proposed network could then be integrated into the vehicles existing perception system by extracting the ego-body contour data and supplying this to the other algorithms which then ignore the area outside the contour coordinates. The proposed network can run at set intervals to save computing power and to check if the camera is misaligned by comparing the new contour data to the previous data.