Fig. 1. Parameters update during training and inference

Fig. 2. Comparison between classification score and IoU overlap score

Fig. 3. Same examples of IoU loss with different locations

Fig. 4. Loss judgement added into centric position

Fig. 5. Network before and after pruning

Fig. 6. Flow chart of channel pruning

Fig. 7. Structure diagram of algorithm

Fig. 8. Initialization unit

Fig. 9. Tracking unit

Fig. 10. Intermediate results of tracking unit

Fig. 11. Online update unit

Fig. 12. Function of online update unit

Fig. 13. System structure diagram of Atlas200 DK

Fig. 14. Test results of CarDark sequence

Fig. 15. Test results of CarScale sequence

Fig. 16. Test results of SUV sequence with occlusion

Fig. 17. Test results of SUV sequence after occlusion