Real Time Object Detection using Single Shot Multibox Detector Network for Autonomous Robotic Arm

Abstract

This paper presents a problem of real-time accurate object detection in picking and placing objects using a robotic arm in uncertain conditions of occlusion, inadequate lighting, and change in camera pose where conventional appearance-based approaches are largely ineffective. Different methods have been proposed to solve these challenges of real time speed and detection under uncertain conditions. Unfortunately, most of these methods have not succeeded in achieving real time speed and detection under uncertain conditions which render them unusable in high performance applications where pick and place robots are used. A robotic pick and place arm that uses a modified version of single shot multibox detector (SSD) convolutional neural network for object detection was developed to tackle this problem. The original SSD network had been used for object detection in other fields like computer science and achieved real time detection but was not accurate in detection especially when it came to small objects and when exposed to uncertainties like varied lighting, occlusion and change in camera pose. The modified SSD network presented in this paper achieved real time detection in the range of 40 frames per second (fps) with accuracies of above 0.75 mAP (mean average precision) in varied lighting conditions, partial occlusion, and changing camera pose. The picking and placing accuracy achieved was also above 80\%. Experimental results validated the performance of the network and robot control method in a realistic scenario of picking and placing objects.