In this study, we propose an algorithm based on machine vision to detect heel-strike events for solving the problem that the gait recognition technology based on wearable sensors is considerably dependent on the cooperation of participants, with high energy consumption and harsh application conditions. The proposed algorithm can accurately detect heel-strike events using ordinary cameras without the cooperation of participants. Initially, we develop an innovative feature for representing gait patterns by designing a consecutive-silhouette difference map (CSD-map). A CSD-map can encode the binary image of several consecutive pedestrian contours extracted from the video frames and output the combination as a single feature map. Different numbers of consecutive pedestrian contour differences result in different types of CSD-map. Further, a convolutional neural network is used for feature extraction and classification of the imaged heel-strike events. In a public database of video data for training and testing, we find 124 individuals under five angles in different wear conditions, and the experimental results obtained using these images denote the accuracy of our method. The identification accuracy is observed to be greater than 93%.