Self-Driving Cars: Evaluation of Deep Learning Techniques for Object Detection in Different Driving Conditions

Deep Learning has revolutionized Computer Vision, and it is the core technology behind capabilities of a self-driving car. Convolutional Neural Networks (CNNs) are at the heart of this deep learning revolution for improving the task of object detection. A number of successful object detection systems have been proposed in recent years that are based on CNNs. In this paper, an empirical evaluation of three recent meta-architectures: SSD (Single Shot multi-box Detector), R-CNN (Region-based CNN) and R-FCN (Region-based Fully Convolutional Networks) was conducted to measure how fast and accurate they are in identifying objects on the road, such as vehicles, pedestrians, traffic lights, etc. under four different driving conditions: day, night, rainy and snowy for four different image widths: 150, 300, 450 and 600. The results show that region-based algorithms have higher accuracy with lower inference times for all driving conditions and image sizes. The second principle contribution of this paper is to show that Transfer Learning can be an effective paradigm in improving the performance of these pre-trained networks. With Transfer Learning, we were able to improve the accuracy for rainy and night conditions and achieve less than 2 seconds per image inference time for all conditions, which outperformed the pre-trained networks