In 2021, companies in logistics and parcel delivery face high-cost pressure and prioritize innovation to save costs or increase revenues in order to stay competitive. In the following, we focus a case study featuring a computer vision based aftermarket collision avoidance system.
The deep learning system is based on viso.ai, a cloud platform for AI vision applications that enables businesses to build, manage and operate their own AI vision applications. For collision avoidance, we leverage the latest AI hardware technologies such as AI chips specialized for video processing (Vision Processing Unit, VPU) and take advantage of edge computing capabilities in the resource-constrained environment of a delivery van.
The case study was developed together with Intel and a worldwide leading postal organization to digitize delivery logistics and collision avoidance to prevent accidents with deep learning and computer vision.
The Need for Collision Avoidance Systems
The global parcel delivery market is growing at an unprecedented speed due to online sales; as a result, traditional postal companies face enormous cost pressure and must take action to stay competitive.
Delivery vans make up a significant share of the postal vehicle fleets. Delivery vans are a key cost driver and their maintenance is costly, primarily because of inefficiencies, disconnected systems, collision accidents, property damage, and personal injuries.
To manage the rapidly growing number of dispatched parcels, the fleets are generally expected to grow in accordance. At the same time, delivery vans are the direct touch point to the end client, hence crucial for the company’s reputation and perceived product quality.
Serious cost-related problems to be addressed include:
- High costs due to accidents and damage
As time pressure increases with a growing number of packages out for delivery, the number of accidents and associated direct costs for repair are at an all-time high. Additionally, van damage has a direct impact on the company’s reputation – especially if an accident causes bodily harm.
- High costs due to time-consuming processes
Along with accidents comes time-consuming processes such as legal fees, indirect costs to fill in damage reports, or arranging replacement vehicles to avoid interruption to delivery service. Moreover, current ways of reporting accidents often face media discontinuity (e.g. fill in forms by hand and on paper, copy it into another system).
- High costs due to parcel theft
Parcel delivery vans are a popular target for theft. Urban areas face increasing numbers of stolen packages, mainly electronic goods. While postal organizations try to control this, there’s still no effective way to prevent and solve these crimes. Often, the only way is to increase staffing for certain routes, hence driving up costs and further impacting profitability.
A multi-functional collision avoidance system for trucks in delivery logistics will help to solve these challenges.
Challenges of In-vehicle Computer Vision
The environment in delivery vans is harsh, causing technical challenges to implementing a vision-based system:
- High AI processing power required
Vision-based applications are resource-intensive. Especially the real-time video processing of multiple streams in parallel requires powerful computing hardware. In-vehicle applications of computer vision are especially difficult because of the space constraints that affect cooling and require a small form factor.
- Power and Vibration Constraints
Van batteries are exposed to continuous stop-and-start activity and are easily depleted, so additional device power consumption must be minimized. The hardware is exposed to continuous vibration inside the delivery van, so it must be vibration-proof.
- Scalability and Cost Effectiveness
As the industry is facing financial pressures and the system needs to be scalable to thousands of vans, it must come at a cost low enough to achieve a return on investment. Therefore, expensive technologies or non-modular systems cannot be used. The total cost of implementation matters as well as the possibility to dynamically replace or upgrade single parts of the system.
Edge Computing Enables Near Real-Time AI and Analytics
Today, many transportation providers rely on disaggregated data platforms and independent point solutions. Logistics companies seek to connect the fleet with the environment, turn data into insights, achieving fast, efficient, and informed use of logistics systems.
Viso.ai technologies enable AI and computer vision analytics in near-real-time, helping support public safety or digitize business processes. With edge computing and inference, logistics companies benefit from fast response times, free up bandwidth, and help keep sensitive data private.
Features of the next-gen Aftermarket Collision Avoidance System
While vehicles become smarter, operations often cannot benefit from automotive systems because of sensitive business information, certification and development processes or operative challenges (e.g. leased fleet, changing manufacturer).
A vision-based solution, based on the viso.ai platform powered by Intel AI technologies such as the Intel Neural Compute Stick 2, uses optical sensors around delivery vans to connect the vehicles with the environment, hence automating last-mile delivery processes. The business automation system is independent of any van manufacturer and can be equipped at or after van procurement in less than two hours.
As the vans are now connected to the cloud and have computing resources locally available, the system is easily able to integrate future innovations. The initial product is focusing on decreasing the costs related to delivery vans by providing the following main functionalities:
- Accident Prevention
The van driver gets a 360° view of the delivery van without any blind spot on a tablet and can easily access every single stream. Compared to automotive-grade systems, the cameras are mounted at optimized positions (above the surroundings) and provide high-resolution streams.
- People Safeguard functionality
For accident prevention, the deep learning model detects people around the delivery van and warns the driver in real-time about potential obstacles.
- Automated workflows
All workflows related to damage management are digitized and can be completed directly on the tablet before the report is synced on a cloud-based damage management portal.
- Theft detection
The system detects anomalies around the van, flags the situation, and starts recording automatically. If an incident happens, the video file can be exported.
The Use Case Setup
Viso.ai software leverages the latest hardware and software technology from Intel to meet these diverse and challenging requirements. The use case runs on Intel Core processors, combined with the Intel Movidius Myriad X VPU for Deep Learning inference, in a robust and industrial housing. Additionally, the Deep Learning models are optimized using the Intel Distribution of OpenVINO Toolkit, a software kit that helps developers and data scientists speed up computer vision workloads, streamline deep learning inference and deployments, and enable execution across a range of Intel’s AI hardware.
The computer is mounted inside the delivery vans. The Movidius Myriad X Vision Processing Unit provides enough power to perform object detection on three input camera streams in real-time at low electrical power consumption. The output is used to warn the driver about potential obstacles, while the Intel Core i3 processor performs geometrical transformation tasks to display the live video streams to the driver inside the delivery van.
The use of the latest, optimized Intel AI technologies has resulted in significant damage-related cost reduction, according to initial KPI assessments after the first month of system usage. Further features are now planned to be integrated into the same system, in order to scale the benefits observed even more.
According to Arkadiusz Hruszowiec, Intel Business Development Manager for the region, “this project is a proof-point that our strategy of enabling AI on edge devices works in practice. By using the Intel Distribution of OpenVINO toolkit and Intel Vision Accelerator Solutions, viso.ai is able to take video streams, analyze them in a deep-learning model and draw insights in real-time at the edge, which results in real business impact for the end customer.”
- See more use cases and computer vision applications
- Learn how viso.ai built a platform for businesses to deliver computer vision
- Read about the challenges of computer vision implementation