Fig. 1. Overview framework of steel scrap grading. (a) Unloading images; (b) keyframes, assigned IDs, and filtered IDs from top to bottom; (c) scrap recognition result; (d) scrap proportion statistics based on IDs and recognition results

Fig. 2. Illustration of bidirectional association algorithm (the hatched area denotes the intra-vehicle region of interest identified via change detection). (a) Backward association method associates objects from the current to the previous frame (the association on the left failed while that on the right succeeded); (b) forward association method successfully associates objects from the previous to the current frame

Fig. 3. Detailed architectures of models for object association. (a) Detailed architecture of object segmenter; (b) detailed architecture of object tracker

Fig. 4. Illustration of backward association algorithm

Fig. 5. Illustration of forward association algorithm (shaded area indicates detected change region)

Fig. 6. AOI-based change detection method. (a) Method of Wang et al. failed in steel scrap scenario; (b) proposed method focuses on both the region of interest and time-series difference, resulting in better robustness

Fig. 7. Proposed steel scrap grating machine vision system. (a) Actual working environment; (b) movable machine vision system

Fig. 8. Data analysis of steel scrap dataset. (a) Change detection data of type 1; (b) change detection data of type 2; (c) box plot of grab interval time; (d) visualization of typical steel scrap images

Fig. 9. Robust analysis of AOI module (circle highlights smoke interference and mask areas indicate the detected change regions). (a)(b) Two temporal keyframe images; (c) change detection result without AOI module; (d) change detection result with AOI module

Fig. 10. Effects of backward and forward association algorithms (different colors indicate objects with different IDs). (a) Input keyframe images; (b) results of object segmenter; (c) results of backward association algorithm only; (d) results of forward association algorithm only, where white areas indicate no objects being detected; (e) association results of our algorithm

Fig. 11. Practical application results of time-series deduplication method for steel scrap (different colors indicate objects with different IDs). (a) Input keyframe images; (b) detection results of our method; (c) results covered with change detection masks