Abstract
A complex area of computer vision is real-time object detection and tracking. Taking hardware and cost into consideration, an affordable, compact, less power-hungry, and high-performance machine Raspberry Pi is used for experiments, which may be used more than ever in near future. In this project, Faster R-CNN and SSD object detection pre-trained models are used. These pre-trained models are fine-tuned and trained for a custom object by using transfer learning and deployed on Raspberry Pi to evaluate if it is suitable for real time object detection. There is a tradeoff between accuracy and speed, where a higher throughput of 22 FPS for SSD is obtained whereas faster RCNN has 9 FPS but comparatively higher accuracy. For tracking the object, a centroid tracking algorithm is implemented. To keep track of objects being entered or removed from a location, a counter is maintained with a threshold line in the frame that counts the object. If it passes the threshold in the desired direction that is down or up, the counter is increased accordingly.