Human Action Recognition (HAR) is a significant area of research in video processing that has gained widespread attention due to its various application domains. Since a single human action occurs in several video frames, spatial-temporal features are the key to HAR. This paper presents a hybrid model for recognizing human actions by capturing spatial and temporal features. The proposed model utilizes a residual deep Convolutional Neural Network (CNN) to collect discriminative spatial information from videos. Then average pooling is applied to downsample the spatial dimension of the feature maps and focus on the most informative features. The squeeze and excitation mechanisms are integrated with the proposed model to highlight important action features. These crucial spatial features are then fed to the Gated Recurrent Unit (GRU) to capture temporal dynamics and the sequential dependencies inherent in human actions over time. Finally, a softmax layer is applied for classifying human actions based on the learned spatio-temporal dynamics. The performance of the presented technique was evaluated on the UCF Sports Action and UT-Interaction datasets and achieved an accuracy of 98.89% and 99.28%, respectively. The experimental findings demonstrate the competitive effectiveness of the presented technique compared to cutting-edge techniques.