Classification of Objects Detected by the Camera based on Convolutional Neural Network

Nowadays, we are trying to achieve as much vehicle autonomy as possible by developing Advanced Driver-Assistance Systems (ADAS). For such a system to make decisions, it should have insight into the environment of the vehicle, e.g. the objects surrounding the vehicle. During forward driving, the information about the objects in front of the vehicle is usually obtained by a front view in-vehicle camera. This paper describes the image classification method of the objects in the front of the vehicle based on deep convolutional neural networks (CNN). Such CNN is supposed to be implemented in embedded system of an autonomous vehicle and the inference should satisfy real-time constraints. This means that the CNN should be structured to have fast inference by reducing the number of operations as much as possible, but still having satisfying accuracy. This can be achieved by reducing the number of parameters which also means that the resulting network has lower memory requirements. This paper describes the process of realizing such a network, from image dataset development up to the CNN structuring and training. The proposed CNN is compared to the state-of-the-art deep neural network in terms of classification accuracy, inference speed and memory requirements.