Human activity recognition system (HARS) is critical for monitoring individuals' health and movements, especially the elderly and the disabled, in different environments. Wearable sensors are valuable tools for information acquisition due to the mobility of people. Researchers have proposed different methods for sensor-based HAR, which face challenges. This paper uses deep learning (DL), which combines the deep convolution neural network (CNN) and long short-term memory (DeepConvLSTM) to extract and select features to recognize high-performance activity. Then, the Softmax-Support Vector Machine (SofSVM) performs the classification and recognition operations. This paper uses a comprehensive nested pipe and filter (NPF) architecture. Filters are usually independent, but some filters can be bidirectional to improve the performance of complex activity recognition. There is two-way communication between convolutional layers and long short-term memory (LSTM). The Opportunity dataset is public and includes complex activities. The results with this dataset display that the proposed work improves the weighting F-measure of HAR. This paper has also shown other experiments; that involve comparing the number of sensors, max-pooling size, and convolutional filter size. According to this dataset, the proposed method weighting F-measure is 0.929. The proposed approach is indeed effective in activity recognition, and the NPF architecture covers all the components of the activity recognition process.