Falls have always been one of the major threats to the health and well-being of elderly people, particularly for those living alone. Both wearable and non-wearable fall detection systems have already been developed. However, the fall detection systems using WiFi channel state information (CSI) have attracted a significant interest from researchers due to their non-intrusive and low-cost nature. There are existing machine learning (ML) based fall detection systems using WiFi CSI; however, most systems trained with comprehensive datasets tend to achieve relatively lower accuracy compared to that of the systems trained with less inclusive datasets. To address these issues, we propose a novel, deep learning based fall detection technique. First, we implement different WiFi CSI collection tools and evaluate their potential for fall detection. To develop a highly accurate fall detection technique, we construct a comprehensive dataset, which consists of over 700 CSI samples including different types of falls and other daily activities, performed in four different indoor environments on and off the dominant paths. With this dataset, we then develop a deep learning based classifier using an image classification algorithm. The proposed technique, unlike the other fall detection systems, only requires down sampling and reshaping in pre-processing. The proposed fall detection system is evaluated with the constructed dataset, and it outperforms two other existing systems. It achieves over 96% accuracy for CSI collected in all four environments and 99% accuracy for CSI collected in certain combinations of the environments.