Improving multi-object detection and tracking with deep learning, DeepSORT, and frame cancellation techniques

Abstract

Multi-object detection and tracking is a crucial and extensively researched field in image processing and computer vision. It involves predicting complete tracklets for many objects in a video clip concurrently. This article uses the frame cancellation technique to reduce the computation time required for deep learning and DeepSORT (for any version of the YOLO detector) coupled with DeepSORT algorithm techniques. This novel technique implements a different number of frame cancellations, starting from one frame and continuing until nine frame cancellations, tabling the result of each frame cancellation against the overall system performance for each frame cancellation. The proposed method worked very well; there was a small drop in the average tracking accuracy after the third frame rate cancellation, but the execution time was much faster.

Keywords

• kalman filter
• multi-object detection
• multi-object tracking
• yolo5 deep learning
• data association metric