Abstract The miniature sensor devices and power-efficient Body Area Networks (BANs) for Human Activity Recognition (HAR) have gained increasing interest in different fields, including Daily Life Assistants (DLAs), medical treatment, sports analysis, etc. The HAR systems normally collect data with wearable sensors and implement the computational tasks with a host machine, where real-time transmission and processing of sensor data raise a challenge for both the network and the host machine. This investigation focuses on the hardware/software co-design for optimized sensing and computing of wearable HAR sensor networks. The contributions include (1) design of a miniature wearable sensor node integrating a Micro-Electro-Mechanical System Inertial Measurement Unit (MEMS IMU) with a Bluetooth Low Energy (BLE) in-built Micro-Control Unit (MCU) for unobtrusive wearable sensing; (2) task-centric optimization of the computation by shifting data pre-processing and feature extraction to sensor nodes for in-situ computing, which reduces data transmission and relieves the load of the host machine; (3) optimization and evaluation of classification algorithms Particle Swarm Optimization-based Support Vector Machine (PSO-SVM) and Cross Validation-based K-Nearest Neighbors (CV-KNN) for HAR with the presented techniques. Finally, experimental studies were conducted with two sensor nodes worn on the wrist and elbow to verify the effectiveness of the recognition of 10 virtual handwriting activities, where 10 recruited participants each repeated an activity 5 times. The results demonstrate that the proposed system can implement HAR tasks effectively with an accuracy of 99.20 %.