Embedded smart cameras are stand-alone units that combine sensing and processing on a single embedded platform. They allow flexibility in camera placement, and provide mobility without being dependent on wired links. On the other hand, detecting and tracking objects from videos captured by mobile cameras is a very challenging task even on powerful computers. Since embedded smart cameras have very limited processing power, memory and energy, the challenge becomes even bigger. In this paper, we present a person detection system using an embedded smart camera mounted on a remote-controlled car. We employ histogram of oriented gradients (HOG) for detection, and present the performance results obtained on these resource-constrained environments. The example application is patrolling hallways in a building to detect people.