Learning to Track Object Position through Occlusion