Real-time shape and pedestrian detection with FPGA

Detecting objects according to their shapes from infrared images is needed in many applications. In this paper, two methods of shape detection are designed and implemented on an FPGA (Field Programmable Gate Array) to achieve real-time performance with embedded system. The first method is for rigid object detection and is named Shape Constraint (SC), which uses 56 binary templates at each scale to represent an object with different viewpoints and rotation angles. The templates are arranged in a tree structure with common regions extracted as nodes on different levels. The computation is completely pipelined in the FPGA. Three toys are used for experiment and can be detected simultaneously from cluttered scenes, thus demonstrating the effectiveness of this method. The second method is for nonrigid objects and is based on Naïve Bayes. Implemented in a pipeline, it enables the FPGA to perform pedestrian detection in real time from infrared images. In contrast to conventional methods in which a huge number of negative samples are collected for training, an even distribution is assumed in our method and no negative samples are needed, thus greatly shortening the training time. The hardware computation architecture of these two methods can be easily applied to video frames of arbitrary resolution, generating detection results for each frame at the same rate as image capturing.