Abstract: A general scheme for automated classification of gait patterns based on process neural networks is proposed to overcome limitations of the feature vector approach, including difficulty in feature selection, heavy computing load, and complexity of recognition algorithms. Initial human gait capture was done using accelerometers mounted on the lower limbs of a subject to obtain kinematic information for various gait patterns. Using a Butterworth filter, a time-series acceleration signal was processed, then fitted as a time-dependent function and presented to PNNs directly. For the capacity of arbitrary functional approximation of PNNs, the automated classification of time-series acceleration signals for different gait patterns was achieved. Furthermore, to overcome the slow convergence and localized minimum problems with the traditional gradient descent method, a particle swarm optimization algorithm was adopted to modify the weight functions and weight values of PNNs. Experimental results demonstrated the correctness and effectiveness of the proposed scheme.