As assistive, wearable robotic devices are being developed to physically assist their users, it has become crucial to develop safe, reliable methods to coordinate the device with the intentions and motions of the wearer. This dissertation investigates the recognition of user intent during flexion and extension of the human torso in the sagittal plane to be used for control of an assistive exoskeleton for the human torso. A multi-sensor intent recognition approach is developed that combines information from surface electromyogram (sEMG) signals from the user’s muscles and inertial sensors mounted on the user’s body. Intent recognition is implemented by following a pattern classification approach, wherein a linear discriminant analysis (LDA) based method of pattern classification is utilized. This method of classification builds on a traditional LDA by utilizing multiple classifiers from multiple sensors that are combined together using a majority voting based classifier fusion scheme, to deliver improved classification performance. Additionally, there is a focus on identification of suitable features for classification. Extraction of features in the time, frequency and time-frequency domains is discussed. Wavelet transform methods are employed for targeted extraction of nonlinear time-frequency domain features, and the effectiveness of these features in improving classification performance is emphasized. Experimental results using sEMG and inertial signals recorded from human subjects, to evaluate the pattern classification and feature extraction methods are presented. Results show that a combined sensor approach that utilizes both inertial and sEMG data leads to a 70% improvement in classification performance. Results also show that the use of multiple time-frequency domain features in conjunction with majority voting based classifier-fusion leads to an additional 75% improvement in classification performance, with a best case of up to 97% accuracy in recognizing user intent. This research has provided an effective demonstration of leveraging nonlinear time-frequency domain features with linear methods of classification to deliver accurate and computationally efficient intent recognition. In addition, the research effort has also developed a library of features that can serve as a starting point for future efforts in classifying torso motions.