Traffic flow information detection under resource limitations based on framework optimization

Most studies found that stacking models improved computational efficiency slightly. These methods are not feasible for most platforms. Optimizing the tracking framework for resource-constrained environments is crucial. In multi-object tracking, detection is the main method. This paradigm has low coupling because components are designed separately. This article introduces a tracking-by-detection framework for traffic flow information detection that optimizes matching. This framework will introduce the Selected Intersection over Union (SIoU) loss function, Detection Error Suppression Module (DESM), and Multi-Detection Error Suppression Module (MDESM). This framework was implemented at Malaysian road intersections for 11 ablation experiments to verify each component's excellent performance. This article improved the Simple Online and Realtime Tracking (SORT) tracking method using SIoU, DESM, and MDESM. The new method tracks 10% better than SORT and 7% better than DeepSORT. It also classifies traffic flow 4% more accurately than SORT. The number of ID switches is only 25% of what SORT and DeepSORT have. Overall, the new method works better than both trackers, even with limited resources.