Object detection from onboard vehicle camera

The purpose of this thesis is threefold. First, DEtection TRansformer (DETR) network performance is evaluated against Single shot detector Resnet50 FPN (SSD) on an ensemble of common open source datasets. Based on this baseline work, a novel algorithm for detecting "interesting" events is developed and evaluated. It is based on the combination of two different object detection models trained on the same data. On one hand a standard object detection model trained to recognize 10 different labels. On the other hand, a smaller mobile friendly network trained to recognize if a given object is there or not, called No Label Network (NOLAN). The standard object detection model is the SSD model outlined above, while the mobile friendly network is the SSD MobileNet. Secondly, a brand new lightweight Linear encoder decoder (LEDD) object detection model to improve performance on small objects is proposed. Lastly, software routines leveraging "Qualcomm Neural Processing SDK for AI" have been developed to efficiently deploy trained models on qualcomm chips present in the car. This thesis is a joint work between: Politecnico di Torino, Volvo Cars and Zenseact.