Mobilenet-SSDv2: An Improved Object Detection Model for Embedded Systems

Object detection plays an important role in the field of computer vision. Many superior object detection algorithms have been proposed in literature; however, most of them are designed to improve the detection accuracy. As a result, the requirement of reducing computational complexity is usually ignored. To achieve real-time performance, these superior object detectors need to operate with a high-end GPU. In this paper, we introduce a lightweight object detection model, which is developed based on Mobilenet-v2. The proposed real-time object detector can be applied in embedded systems with limited computational resources. This is one of the key features in the design of modern autonomous driving assistance systems (ADAS). Besides, we also integrate a feature pyramid network (FPN) with the proposed object detection model to effectively improve detection accuracy and detection stability. Experimental results show that the proposed lightweight object detection model achieves up to 75.9% mAP in the VOC dataset. Compared with the existing Mobilenet-SSD detector, the detection accuracy of the proposed detector is improved about 3.5%. In addition, when implemented on the Nvidia Jetson AGX Xavier platform, the proposed detector achieves an average of 19 frames per second (FPS) in processing 720p video streams. Therefore, the proposed lightweight object detector has great application prospects.