Object detection in real-time is an important aspect in robotics, surveillance and autonomous systems. Nevertheless, to be able to have speed and accuracy on embedded devices with resource constraints is still a challenge. This study examines the lightweight object detection models made on deep-learning platform, namely YOLOv8-N and MobileNet-SSD, which are going to be deployed on the Raspberry Pi 4 and Raspberry Pi 5 hardware. We compare strategies of optimization, pruning and sensor fusion methods in order to improve detection in dynamic environment. On experimental outcomes, it is shown that lightweight architectures are capable of achieving accuracy and real-time responsiveness, and are, therefore, appropriate to the perception of mobile robots.