Obstacle Detection in Autonomous Vehicles Using Deep Learning

This project was carried out to enhance the intrinsic intelligence and surveillance of an autonomous vehicle (AV) with remote control capability in Robusta project – sponsored by Tekes – the Finnish Funding Agency for Technology and Innovation, done at Helsinki Metropolia UAS. The scope included developing four software used namely in training deep learning models used in object recognition, real-time obstacle recognition, real-time obstacle detection, and object detection. Five models were trained with the model trainer (MT) software using CIFAR-10 dataset. The autonomous vehicle obstacle detector (AVOD) prototype was developed to be used in the AV’s internal intelligence system. Using this prototype, possibly with a lidar: not included in this report, obstacle detection – recognition and localization of driving scene impediment object – could be carried out by the AV. Given it is a real-time application, the software’s processed frames per second (fps) should nearly equal the camera’s fps. An object detector mobile (ODM) Android® application was developed for object detection. Using the trained model obstacle recognizer (TMOR) a model recorded 85.6% prediction accuracy with 10,000 test images. The AVOD software prototype could, in real-time, detect in a frame all objects in its trained classes (21) – processing 30fps at 100.0% the camera’s capacity.The software and training results are on a gitlab private account. It can be oprovided on request with full acknowledgment of the author's work