Author Affiliations
1College of Physics and Information Engineering, Fuzhou University, Fuzhou 350108, China2Department of Electrical and Computer Engineering, Dalhousie University, Halifax, NS B3J 1Z1, Canadashow less
Fig. 1. Infrared array sensor,(a)physical view of the sensor,(b)infrared temperature image(palm)
Fig. 2. System composition design
Fig. 3. Flow chart of the head motion detection algorithm
Fig. 4. Original image and pseudo-color image(the region of human head and shoulder)(a)origin image(32×32),(b)pseudo-color image(32×32),(c)pseudo-color image(64×64)
Fig. 5. Flowchart of the head salient region extraction algorithm
Fig. 6. Comparison of preprocessed results
Fig. 7. 3D image fusion of a sequence of frames
Fig. 8. Residual learning structure of ResNeXt network,(a)BottleNeck structure of ResNet network,(b)split-transform-merge structure of Inception network,(c)block structure of ResNeXt network
Fig. 9. Pro-ResNeXt50 network
Fig. 10. The training accuracy and loss of three network
Fig. 11. Experience scenarios(a)Experiments in an indoor hall,(b)Experiments in a car:Experiments were conducted in an indoor hall to simulate a driving and online learning environment,as shown in Fig. 11(a). The test user is sitting on a chair,and the sensor is fixed at the height of 1.2 m above the ground by a tripod so that it is aligned with the user’s front face. The collection distance ranges of 0.5 m to 1 m,and the collected lighting environment includes both day and night conditions.
Fig. 12. Accuracy using different methods
Fig. 13. Random continuous head movement steering
Fig. 14. Recognition accuracy in different detection distances and light conditions
Algorithm 1Adaptive Threshold |
---|
Input: the image of Canny edge detection: | the local window size | 1. Obtain the threshold(C)value by the OTSU method
| 2. Obtain the imageafter mean filtering:
| 3. Obtain the continuous boundary image()
| Output: the continuous boundary image |
|
Table 0. [in Chinese]
Item | Specification |
---|
Infrared sensor model | HTPA 32×32 | Camera | 1 | Temperature range of object | -40~85℃ | Viewing angle | 66° | Number of pixels | 1024(32×32) | Temperature output mode | | Frame rate | 5 frames/s |
|
Table 1. HTPA infrared sensor specification parameters
Network | ResNet50 | ResNeXt50 | Pro-ResNeXt50 |
---|
#params. | 25.5 | 25.0 | 22 | FLOPs | 4 | 4 | 4 |
|
Table 2. Comparison of the three networks
Activity | Front | Bow | L45° | L90° | LC | LT | R45° | R90° | RC | RT |
---|
Total | 300 | 280 | 159 | 154 | 160 | 170 | 173 | 167 | 171 | 182 | Precision | 0.947 | 0.989 | 0.962 | 0.961 | 0.981 | 0.970 | 0.948 | 0.964 | 0.982 | 0.962 |
|
Table 3. The Precision for each activity
Method | Original+ Pro-ResNeXt50 | Original +CBAM +ResNeXt50 | Channel(1)+ Pro-ResNeXt50 | Channel(1,2)+ Pro-ResNeXt50 | Channel(1,3)+ Pro-ResNeXt50 | Channel(1,2,3)+ Pro-ResNeXt50 |
---|
Accuracy | 87.73% | 94.47% | 89.35% | 92.06% | 87.31% | 96.76% |
|
Table 4. Accuracy using different channels
Method | Accuracy | Times |
---|
ResNet50 | 94.10% | 7f/s | ResNeXt50 | 94.83% | 8f/s | Pro-ResNeXt50 | 96.76% | 9f/s |
|
Table 5. Accuracy and time using different network
Activity | Front | Bow | L90° | R45° | RC | RT | Average Acc |
---|
Total | 267 | 184 | 205 | 195 | 189 | 192 | — | Precision | 0.948 | 0.902 | 0.942 | 0.910 | 0.963 | 0.958 | 0.937 |
|
Table 6. The Precision for each activity in the car