An Embedded Implementation of Improved SSD with RTMaps and NXP Bluebox2.0 for Autonomous Platforms

Abstract

Multi target object detection is predominantly employed in field of computer vision. Computer vision and machine vision goes hand in hand to develop efficient applications such as autonomous driving systems and intelligent robotic systems in industries. One of the widely used neural network architectures for object detection is Single Shot MultiBox Detector (SSD). Numerous approaches have been carried out in this field by introducing complex convolutional operations to enhance performance. Still it remains an onerous challenge to develop an efficient neural network model can detect small objects accurately. To overcome this challenge, this research proposes an improved version of SSD enhancing the performance of auxiliary convolutional layers for small object detection. Dice coefficient and cross entropy loss approach was adapted to calculate Intersection Over Union (IoU) threshold. The model was trained on PASCAL VOC 2007 and 2012 datasets and testing was carried on PASCAL VOC 2007. Mean average precision (mAP) of 80.7%, 3.2% greater than original SSD at an input size of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$300 \times 300$</tex> . The architecture model was deployed in Intempora RTMaps embedded with NXP Bluebox 2.0 to observe the model performance in real-time.