IIC Journal of Innovation 15th Edition | Page 39

Safely Back to Work
have a balanced dataset , i . e ., an almost equal number of training examples from both the classes , so that the model ’ s performance is not biased towards any particular class 22 .
As an example , we trained a MobileNetV2 classifier 23 on 750 images each of masked and unmasked human faces , with an 80 % -20 % split between training and testing images 24 . With this model , we were able to achieve a validation accuracy of ~ 97 % and F1 score > 0.97 for both the classes . Another major aspect is the speed of inference of the model as an ideal monitoring application needs to be real-time . For good speed , it needs to be ensured that the model is lightweight , i . e ., computationally less expensive . One way to achieve this is by using convolutional neural network architectures that use depth-wise separable convolution rather than simple convolution 25 . This would reduce the computation load and increase the speed of inference of the system .
LiDAR
While a great technology with a good range and resolution between objects , LiDAR requires detailed planning , as an entire view of the space is necessary for more granular tracking . Therefore , it is recommended to mount LiDAR systems at least 5-6 feet above the ground in order to get a 360-degree view of the space and be fully visible within the space 26 . Like most light spectrum-based tracking systems , LiDAR is not able to penetrate walls and requires a nonobstructed view of the objects that need to be tracked . For this reason , LiDAR systems should be placed in every enclosed room or above smaller separated areas such as cubicles or desks . Though LiDAR systems can scan up to a maximum range of 200 meters ; the accuracy in tracking people is best at a range of 50 meters 27 . Given that LiDAR can work in various types of weather , use cases can be expanded to outdoor solutions that may need to track the number of people in a public space , and / or a parking lot . This is especially useful for social distancing use cases where social distancing is necessary in both indoor and outdoor spaces .
22 https :// machinelearningmastery . com / what-is-imbalanced-classification /
23 https :// towardsdatascience . com / review-mobilenetv2-light-weight-model-image-classification-8febb490e61c
24 https :// machinelearningmastery . com / train-test-split-for-evaluating-machine-learning-algorithms /
25 https :// www . geeksforgeeks . org / depth-wise-separable-convolutional-neural-networks /
26 http :// teecom . com / media / TEECOM-Occupancy-Measurement-Analysis-Report . pdf
27 https :// www . konicaminolta . com / us-en / future / 3dlr / index . html
- 34 - November 2020