Author Affiliations
College of Physics and Electronic Engineering, Shanxi University, Taiyuan, Shanxi 030006, Chinashow less
Fig. 1. YOLOv3 network structure diagram. (a) Overall structure diagram of YOLOv3 network; (b) structure diagram of set conv layer and YOLO layer
Fig. 2. Dataset analysis results. (a) Target width and height distribution of the dataset; (b) k-means clustering analysis result
Fig. 3. Average IOU. (a) Relationship between Avg IOU and Mean IOU; (b) Avg IOU of k-means and k-thresh
Fig. 4. Image of object detection
Fig. 5. Video-YOLOv3 network structure. (a) Overall structure diagram of video-YOLOv3 network; (b) structure diagram of splice-conv module
Fig. 6. Comparison of network structure. (a) YOLOv3 network structure; (b) video-YOLOv3 network structure
Fig. 7. Flow chart of predicting new image
Fig. 8. Loss function and Avg IOU curve of video-YOLOv3. (a) Loss function curve; (b) Avg IOU curve
Fig. 9. Comparison of test results. (a) YOLOv3 test results; (b) video-YOLOv3 test results
Fig. 10. Comparison of real-time detection. (a) Detection every 5 frames; (b) detection every 6 frames; (c) detection every 7 frames
Class | Person | Tvmonitor | Chair |
---|
Total number | 2503 | 1050 | 1958 |
|
Table 1. Number of images with each type of object in the dataset
Method | AP | mAP |
---|
Person | Tvmonitor | Chair |
---|
Faster R-CNN | 70 | 74 | 65 | 69.67 | SSD | 72 | 73 | 58 | 67.67 | YOLOv3 | 77 | 74 | 60 | 70.33 | Video-YOLOv3 | 77 | 76 | 64 | 72.33 |
|
Table 2. mAP values of different models on the dataset unit: %
Item | CPU(Intel i7-7700k) | GPU( Tesla K40) |
---|
Original | Corrected | Original | Corrected |
---|
Time /ms | 231.22 | 51.78 | 67.54 | 16.39 | Max /(frame/s) | 4.33 | 19.45 | 15.73 | 64.26 |
|
Table 3. Comparison of CPU and GPU real-time detection time