研究目的
To develop a dynamic hand gesture recognition system for driving assistance using FMCW radar and CNN to classify nine hand gestures and explore the impact of different experimental scenarios on accuracy.
研究成果
The hand gesture recognition system using 77GHz FMCW radar and CNN achieves an average accuracy of 94.4% for nine gestures in controlled conditions. Accuracy is affected by distance and angle variations, with some gestures (e.g., opening palm) being more robust. Future work should include more diverse scenarios and users to improve robustness and expand the gesture set.
研究不足
The system's accuracy decreases with increasing distance and angle due to reduced signal-to-noise ratio and changes in micro-Doppler signatures. The CNN was trained on data from a single user in controlled environments, limiting generalizability to different users and scenarios. Gestures with high similarity (e.g., swiping motions) show higher misclassification rates.
1：Experimental Design and Method Selection:
The study uses a 77GHz FMCW radar sensor to capture micro-Doppler signatures of hand gestures, which are processed and classified using a convolutional neural network (CNN) based on the LeNet architecture. Data augmentation techniques such as adjusting sharpening, saturation, and brightness of spectrograms are employed to enhance training data robustness.
2：Sample Selection and Data Sources:
Nine hand gestures (e.g., swiping, opening palm) are performed by a participant at an average radial distance of 20 cm from the radar. Each gesture is measured 40 times, resulting in 360 data pieces per experiment set.
3：List of Experimental Equipment and Materials:
The primary equipment is the IWR1443 77GHz FMCW radar sensor from Texas Instruments, which includes onboard antennas with four receivers and three transmitters. Software tools include MATLAB for signal processing and model training.
4：Experimental Procedures and Operational Workflow:
The radar is fixed on a table, and gestures are performed in front of it. Radar parameters are set (e.g., 255 chirps per frame, 40.96 ms frame period, 160 μs chirp time, 4 GHz bandwidth). Echo signals are received, sampled, and stored. Data is preprocessed through pulse compression and short-time Fourier transform to generate spectrograms. The CNN is trained with 75% of data and tested on the remaining 25%, using 200 epochs and a batch size of
5：96 ms frame period, 160 μs chirp time, 4 GHz bandwidth). Echo signals are received, sampled, and stored. Data is preprocessed through pulse compression and short-time Fourier transform to generate spectrograms. The CNN is trained with 75% of data and tested on the remaining 25%, using 200 epochs and a batch size of Data Analysis Methods:
2.
5. Data Analysis Methods: The CNN model's performance is evaluated using accuracy metrics from confusion matrices. Mean square error during training is monitored, and the impact of variables like distance and angle on accuracy is analyzed statistically.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
