Human Activity Recognition (HAR) is a key enabler of various applications, including smart homes, health care, Internet of Things (IoT), and virtual reality games. A large number of HAR systems are based on wearable sensors and computer vision. However, a challenge that has emerged in the last few years entails recognizing human activities using WiFi Channel State Information (CSI). Exiting state-of-the-art solutions have considered only amplitudes of the CSI to recognize human activities, we explore both amplitudes and phase differences to recognize activities. We utilize Continuous Wavelet Transform (CWT) to generate scalogram images from the CSI measurements. Then, we use these images as input to the pertained Convolution Neural Network (CNN), namely AlexNet to extract features that are resilient to environment changes and classify the activities. The experimental results show that the proposed method achieves an accuracy of 98.18% +/- 1.26% using amplitude and phase difference. We also studied the impact of different environments and people, and the results show its robustness.