A highly efficient computer-vision system that can detect moving vehicles with great accuracy and relative motion, all while consuming minimal power.
By capturing moving vehicle images, aerial cameras can provide information about traffic conditions, which is beneficial for governments and planners to manage traffic and enhance urban mobility. Detecting moving vehicles with low-powered devices is still a challenging task. We are going to tackle this problem using a Brainchip Akida neural network accelerator.
Hardware Selection
In this project, we'll utilize BrainChip’s Akida Development Kit. BrainChip's neuromorphic processor IP uses event-based technology for increased energy efficiency. It allows incremental learning and high-speed inference for various applications, including convolutional neural networks, with exceptional performance and low power consumption.
The kit consists of an Akida PCie board, a Raspberry Pi Compute Module 4 with Wi-Fi and 8 GB RAM, and a Raspberry Pi Compute Module 4 I/O Board. The disassembled kit is shown below.
The Akida PCIe board can be connected to the Raspberry Pi Compute Module 4 IO Board through the PCIe Gen 2 x1 socket available onboard.
Setting up the Development Environment
After powering on the Akida development kit, we need to log in using an SSH connection. The kit comes with Ubuntu 20.04 LTS and Akida PCIe drivers preinstalled. Furthermore, the Raspberry Pi Compute Module 4 Wi-Fi is preconfigured in Access Point (AP) mode.
Completing the setup requires the installation of a few Python packages, which requires an internet connection. This internet connection to the Raspberry Pi 4 can be established through wired LAN. In my situation, I used internet sharing on my Macbook with a USB-C to LAN adapter to connect the Raspberry Pi 4 to my Macbook.
To log in and install packages execute the following commands. The password is brainchip for the user ubuntu.
Capturing video of moving traffic using a drone is not permitted in my area so I used a license-free video from pexels.com (credit: Taryn Elliot). For our demo input images, we extracted every 5th frame from the pexels.com video using the Python script below.
python
import cv2
import sys
if __name__ == '__main__':
video_path = './traffic-3840x2160-24fps.mp4'
videoCapture = cv2.VideoCapture(video_path)
if not videoCapture.isOpened():
logging.error("Cannot open video file")
sys.exit(-1)
i = 0
j = 0
while True:
ret, frame = videoCapture.read()
if ret:
img = frame[440:440+1280, 1280:1280+1280]
cv2.imshow('Frame', img)
i = i + 1
if i % 5 == 0:
j += 1
cv2.imwrite(f'./data/images/img_{j:04d}.jpg', img)
key = cv2.waitKey(1) & 0xFF
if key == ord('q'):
print('Exited gracefully')
cv2.destroyAllWindows()
break
We will use Edge Impulse Studio to build and train our demo model. This requires us to create an account and initiate a new project at https://studio.edgeimpulse.com.
To upload the demo input images extracted from the pexels.com video into the demo Edge Impulse project, we will use the Edge Impulse CLI Uploader. Follow the instructions at the link: https://docs.edgeimpulse.com/docs/cli-installation to install the Edge Impulse CLI on your host computer..
The command above will upload the demo input images to Edge Impulse Studio and split them into "Training" and "Testing" datasets. Once the upload completes, the demo input datasets are visible on the Data Acquisition page within Edge Impulse Studio..
Now we can label the data with bounding boxes in the Labeling queue tab as shown in the GIF below.
Model training
Go to the Impulse Design > Create Impulse page, click Add a processing block, and then choose Image. This preprocesses and normalizes image data, and optionally allows you to choose the color depth. Also, on the same page, click Add a learning block, and choose Object Detection (Images) - BrainChip Akida™ which fine-tunes a pre-trained object detection model specialized for the BrainChip AKD1000 PCIe board. This specialized model permits the use of a 224x224 image size, which is the size we are currently utilizing. Now click on the Save Impulse button.
On the Image page, choose RGB as color depth and click on the Save parameters button. The page will be redirected to the Generate Features page.
Now we can start feature generation by clicking on the Generate features button:
After feature generation, go to the Object Detection page and click on Choose a different model and select Akida FOMO. Then click on the Start training button. It will take a few minutes to complete the training.
The FOMO model uses an architecture similar to a standard image classification model which splits the input image into a grid and runs the equivalent of image classification across all cells in the grid independently in parallel. By default the grid size is 8x8 pixels, which means for a 224x224 image, the output will be 28x28 as shown in the image below.
For localization, it cuts off the last layers of the classification model and replaces this layer with a per-region class probability map, and subsequently applies a custom loss function that forces the network to fully preserve the locality in the final layer. This essentially gives us a heat map of vehicle locations. FOMO works on the constraining assumption that all of the bounding boxes are square, have a fixed size, and the objects are spread over the output grid. In the aerial view images, vehicles look similar in size hence FOMO works quite well.
Confusion Matrix
Once the training is completed we can see the confusion matrices as shown below. By using the post-training quantization, the Convolutional Neural Networks (CNN) are converted to a low-latency and low-power Spiking Neural Network (SNN) for use with the Akida runtime. We can see in the below image, the F1 score of 94% of the Quantized (Akida) model is better than that of the Quantized (int8) model.
Model Testing
On the Model testing page, click on the Classify All button which will initiate model testing with the trained model. The testing accuracy is 100%.
Deployment
We will be using Akida Python SDK to run inferencing, thus we will need to download the Meta TF model (underlined in red color in the image below) from the Edge Impulse Studio's Dashboard. After downloading, copy the ei-object-detection-metatf-model.fbz model file to the Akida development kit using command below.
The application loads the MetaTF model (*fbz) and maps it to the Akida neural processor. The inferencing is done on the images from a video file. We have converted several Edge Impulse C++ SDK functions to Python to preprocess FOMO input. Below is the main code for inferencing.
To run the application, login to the Akida development kit and execute the commands below.
$ git clone https://github.com/metanav/vehicle_detection_brainchip_edge_impulse.git
$ cd vehicle_detection_brainchip_edge_impulse
$ python3 main.py
The inferencing results can be accessed at http://:8080 using a web browser. The application also displays the model summary mapped on the Akida PCIe neural processor on the console.
Notice there is no Softmax layer at the end of the model. That layer has been removed during model conversion to run on the Akida processor. The Softmax operation is performed in the application code, rather than in the model..
Demo
The video used for the demonstration runs at a framerate of 24 fps, and the inferencing takes approximately 40ms per frame, resulting in real-time inferencing.
Conclusion
This project highlights the impressive abilities of the Akida PCIe board. Boasting low power consumption, it could be used as a highly effective device for real-time object detection in various industries for numerous use cases.
