Object detection and tracking with decoupled DeepSORT based on αβ filter

Abstract

With the rapid growth of the population, the demand for autonomous video surveillance systems has substantially increased. Recently, artificial intelligence has played a key role in the development of these systems. In this paper, we present an enhanced autonomous system for object detection and tracking in video streams, tailored for transportation and video surveillance applications. The system comprises two main stages: detection stage; this stage employs you only look once (YOLO)v8m, trained on the KITTI dataset, and is configured to detect only pedestrians and cars. The model achieves an average precision of 97.3% and 87.1% for cars and pedestrians classes respectively, resulting a final mean average precision (mAP) of 92.2%. Tracking stage; the tracking component utilizes the DeepSORT algorithm, which originally incorporates a Kalman filter for motion prediction and performs data association using cosine and Mahalanobis distances to maintain consistent object identifiers across frames. To improve tracking performance, we introduce two key modifications to the original DeepSORT: architecture modification and Kalman filter replacement. The tracking tests are carried out on KITTI and MOTChallenge Benchmarks. The final order tracking accuracy (HOTA) scores achieve 77.645 and 54.019 for Cars and Pedestrians classes respectively in the KITTI-Benchmark and 45.436 for the Pedestrians class in the MOTChallenge-Benchmark.