Object detection in thermal infrared spectrum provides more reliable data\nsource in low-lighting conditions and different weather conditions, as it is\nuseful both in-cabin and outside for pedestrian, animal, and vehicular\ndetection as well as for detecting street-signs & lighting poles. This paper is\nabout exploring and adapting state-of-the-art object detection and classifier\nframework on thermal vision with seven distinct classes for advanced\ndriver-assistance systems (ADAS). The trained network variants on public\ndatasets are validated on test data with three different test approaches which\ninclude test-time with no augmentation, test-time augmentation, and test-time\nwith model ensembling. Additionally, the efficacy of trained networks is tested\non locally gathered novel test-data captured with an uncooled LWIR prototype\nthermal camera in challenging weather and environmental scenarios. The\nperformance analysis of trained models is investigated by computing precision,\nrecall, and mean average precision scores (mAP). Furthermore, the trained model\narchitecture is optimized using TensorRT inference accelerator and deployed on\nresource-constrained edge hardware Nvidia Jetson Nano to explicitly reduce the\ninference time on GPU as well as edge devices for further real-time onboard\ninstallations.\n