2 of 6

Solutions/ Devices	Credit: BSD Blind Spot Detection System (adapted from car), INNOVV.com (bike-specific)
Size	Two sensors of 14 x 6.5 x 2.5 cm, 15 m wires, total 653 grams (for BSD Systems)	8.5 x 8.5 x 5.2 cm, 216 grams	(Android Phone - Vivo Y71 model size as proxy) 15.6x7.6x0.7 cm, 150 grams
Detection accuracy	Radar based - High accuracy, warning on detecting moving vehicles at threatening velocity	Ultrasonic-based- Higher accuracy , warning if only detecting moving vehicles at threatening velocity	Computer-vision to detect accelerating vehicles from behind from adjacent & across the lanes. (currently 70+% confidence)
Warning signal	Light emblem warning	Piezo Electronic Buzzer Alarm 95DB Active device speaker 10b	Red-framing of threats shown on phone with sound (vs Green-framing of non-threats)
Installation	Need re-wirings of motorcycle (OEM warranty issue)	GoPro-based clip-and-go standard kit	Simple app download on mounted phone
Charging	Need wiring to motorcycle battery (OEM warranty issue)	3.6V 3200mAh battery/ USB- input	Normal USB-C phone charging
Cost	~100-290 USD	~12.0 USD	~10 USD (for one-time app download cost ) USD 30+3 (Y71 phone+ Mounting)

Project 2 - Vision Guard Build summary: Overall feature comparison

Sonic Guard V2.0 (Project 1)

Vision Guard V 1.0 (Project 2)

Existing typical market offerings

Completed

Planned

3 of 6

Project 2 Vision Guard Build summary: Coding and Reference

Vision Guard V 1.0

Git hub link	https://github.com/Sxneat/VisionGuard
Coding Overview	AI Module, utilizes YOLO/MiDaS systems for object detection and Roboflow for dataset creation and organization of 1000+ photos captured through a phone camera during real motorcycle rides, ie. processes labeled training data to enhance vehicle recognition accuracy and confidence. Detects and classifies vehicles using AI, ie. places bounding boxes around motorcycles and cars with distinct labels, assigning a unique ID per vehicle to track movement even when frames are obstructed. Calculates object velocity, based on the MiDas distance detection system and x/y coordinate tracking, ie. measures position changes over time to estimate speed and direction of motion. Identifies vehicles in the blind spot, ie. turns the bounding box red and displays text alerts when a nearby object is detected within a critical range and moving at a high velocity, prioritizing potential threats.
Coding references	Source for YOLO modules: https://github.com/ultralytics/ultralytics Source for MiDaS modules: https://github.com/isl-org/MiDaS Source for Roboflow modules: - Platform: https://roboflow.com/ - SDK: https://github.com/roboflow/roboflow-python

4 of 6

Project 2 - Vision Guard Build summary: Simulated Testing Results

Vision Guard V 1.0

Testing Results
Comments/Improvements	The program did well in detecting accelerating vehicles from behind both from the adjacent lanes and far away lanes, while not creating fault alert for the other vehicles that are not reducing distances to the rider Further improvement of the coding are as follows Implement a more precise method for measuring vehicle distance to enhance detection accuracy. Introduce additional variables to refine blind spot threat assessment and reduce false alerts. Conduct further testing to determine specific velocity thresholds that pose a risk. Develop a system to convert program distance values into real-world units for clearer interpretation. Create more warning conditions to account for a wider range of potential driving hazards.

Blind spot from far-away lanes

Blind spot from adjacent lanes

5 of 6

Activities	1	2	3	4	5	6	7	8	9	10
Step 1: Conduct further rider interviews (target 10 riders) to reconfirm user journey, pain points, and needs	X
Step 2: Collect the actual front camera VDO samples from the real riding on the road, categorizing them into several key motorcycle blind-spot scenarios	X	X
Step 3: Program train the computer vision algorithm with the actual VDO clips, targeting more than 70% accuracy of blind-spot-related collision predictions		X	X	X
Step 4: Converting the algorithm/coding into a user-friendly application for either Android platform				X	X	X	X	X	X
Step 5: Install the app into the mounted mobile phone and conduct iterative blind spot testing and algorithm improvement for a motorcycle rider on the simulated testing field (the 16m x 6 m car parking space)						X	X	X	X
Step 6: Summarize results from the prototype testing and define next steps for improvement and scaling.										X

Project 2 - Vision Guard Build summary: Learning and Next Steps

Month

Completed

On-going

Planned

Learning from Step 1-3: Four challenges that I aim to address next time;

1. Car object detection modules failed for motorcycles, requiring a custom solution using depth estimation to determine velocity. Future collaboration with motorcycle-based researchers is essential.

2. We underestimated the data required for M/L, having only 3 clips when at least 50 are needed for verified confidence. Given budget constraints, I will plan for creating synthetic data using game engines.

3. We overlooked the transition to mobile phones, as the depth estimation system is too large for phone processing. I need to explore remote clouds or other alternatives.

4. We encountered human behavior constraints, such as balancing safety versus usability. Ensuring a safety buffer while avoiding alert fatigue with false positives was challenging. It reminds me to better understand the role of human behavior in further AI/App deployment.

6 of 6

Project 2 - Vision Guard Build summary: Research References

1. Aquilina, D., and Gatt, T. “Motorcycle Blind Spot Detection Through Computer Vision Techniques.” 2023 ISPA, Rome, Italy, 2023, pp. 1-6, doi: 10.1109/ISPA58351.2023.10279254.

2. Fernández, C., Llorca, D. F., Sotelo, M. A., et al. “Real-Time Vision-Based Blind Spot Warning System: Experiments with Motorcycles in Daytime/Nighttime Conditions.” Int. J. Automot. Technol., vol. 14, 2013, pp. 113–122. https://doi.org/10.1007/s12239-013-0013-3.

3. Hashim, M. S. M., et al. “Determination of Blind Spot Zone for Motorcycles.” IOP Conf. Ser. Mater. Sci. Eng., vol. 670, IOP Publishing, Nov. 2019, pp. 012075–75, https://doi.org/10.1088/1757-899x/670/1/012075.

4. Shen, Y. “Blind Spot Warning Using Deep Learning.” cerv.aut.ac.nz, accessed 25 Jan. 2025.

5. Singh, N., and Ji, G. “Computer Vision Assisted, Real-Time Blind Spot Detection Based Collision Warning System for Two-Wheelers.” 2021 ICECA, Coimbatore, India, 2021, pp. 1179-1184, doi: 10.1109/ICECA52323.2021.9676158.

6. Zuraimi, M. A. Bin, and Zaman, F. H. K. “Vehicle Detection and Tracking Using YOLO and DeepSORT.” 2021 ISCAIE, Penang, Malaysia, 2021, pp. 23-29, doi: 10.1109/ISCAIE51753.2021.9431784.

1 of 6

2 of 6

3 of 6

4 of 6

5 of 6

6 of 6