This paper presents an FPGA-based system for detecting people from video. The system is designed to use JPEG-compressed frames from a network camera. Unlike previous approaches that use techniques such as background subtraction and motion detection, we use a machine-learning-based approach to train an accurate detector. We address the hardware design challenges involved in implementing such a detector, along with JPEG decompression, on an FPGA. We also present an algorithm that efficiently combines JPEG decompression with the detection process. This algorithm carries out the inverse DCT step of JPEG decompression only partially. Therefore, it is computationally more efficient and simpler to implement, and it takes up less space on the chip than the full inverse DCT algorithm. The system is demonstrated on an automated video surveillance application and the performance of both hardware and software implementations is analyzed. The results show that the system can detect people accurately at a rate of about 2.5 frames per second on a Virtex-II 2V1000 using a MicroBlaze processor running at 75   MHz , communicating with dedicated hardware over FSL links.