Fig. 1. Weld feature points of different joint types. (a) Lap weld; (b) fillet weld; (c) butt weld

Fig. 2. Framework diagram of the improved target tracking algorithm

Fig. 3. Targets and corresponding filter response outputs. (a) Ideal response output; (b) filter response output when the target is heavily disturbed

Fig. 4. Response distributions of image block. (a) Two-dimensional response diagram; (b) three-dimensional response diagram

Fig. 5. Partical images for algorithm evaluation

Fig. 6. Typical states during the identification processes of different algorithms. (a) Proposed method; (b) KCF; (c) TLD; (d) MIL; (e) Boosting; (f) MOSSE

Fig. 7. Evaluation results of proposed algorithm. (a) Variation in APCE value throughout the test and their situations at valley values; (b) recognition results of proposed method under valley values of APCE

Fig. 8. Comprehensive evaluation results of performances for different algorithms. (a) Accuracies of different algorithms; (b) time consumptions of different algorithms

Fig. 9. Experimental applications of proposed recognition algorithm in actual weld detection. (a) Actual workspace of the robot; (b) on-site assembly of a welded section in the workspace; (c) assembled weld workpiece; (d) robot reaches the starting point of the weld; (e) robot starts welding at the starting point; (f) an intermediate process of robot automatic welding; (g) robot stops the arc at the end of the weld