VoNet: Vehicle orientation classification using convolutional neural network

This paper presents a novel convolution neural network for classifying the orientation (or viewpoint) of a vehicle in a given image. Current equipping sensors in self-driving car is able to produce bounding box of vehicles in the proximity, but it does not recognize the viewpoint of them. Analyzing surrounding cars' direction in very complex environment has a significant role for autonomous driving. Utilizing nothing but a captured image, the purpose of this research is to classify viewpoint of vehicle: (1) front; (2) rear; (3) side; (4) front-side; and (5) rear-side. Deep convolutional neural network is used as the tool in performing classification task. The approach involves examining different CNN architectures using a large scale car dataset. In addition to that, the goal of the model is to be small and fast enough for limited hardware resource. We are able to achieve 95% accuracy, 57ms inference time on NVIDIA GRID K520 GPU, and 1.6 MB Caffe model size.