ICASSP 2027 Submission

LiteBallNet: A Lightweight Multi-Frame Ball Detector for Cross-Sport Edge Deployment

Mingqi Chen1, Zhaofeng An1, Ruixin Li1, Yun Luo1, Qiang Li1,*, Peizheng Wang1,*

1 School of Artificial Intelligence, Shenzhen Technology University, Shenzhen, China

* Corresponding authors: liqiang1@sztu.edu.cn and wangpeizheng@sztu.edu.cn

Supported by the Natural Science Foundation of Top Talent of SZTU (Grant No. GDRC202411).

45,644parameters
1.876GMAC
86.30FPS · RTX 3060 Laptop
37.0 msRK3588 RKNN · ≈30 FPS
Paper overview

Abstract

Vision-based tiny-ball tracking underpins automated officiating and fine-grained tactical analysis. However, deployment on embedded platforms remains difficult: lightweight networks often lose spatial detail, whereas high-capacity trackers exceed practical memory and computation budgets. To reconcile localization accuracy with edge efficiency, we propose LiteBallNet, a unified multi-frame heatmap network for cross-sport tiny-ball detection. Depthwise separable convolutions reduce computation, limited spatial reduction and U-Net skip fusion preserve fine-grained target cues, and bottleneck Coordinate Attention strengthens direction-sensitive localization. We further introduce structured hard-negative supervision to suppress recurrent court-like heatmap peaks during training without adding inference overhead. Across three public benchmarks, LiteBallNet achieves state-of-the-art F1-score on the evaluated table-tennis benchmark and competitive accuracy on badminton and tennis with only 45,644 parameters. It also sustains real-time inference on both a desktop GPU and an RKNN-deployed edge NPU, establishing a strong accuracy–capacity operating point for portable cross-sport ball tracking.

01

A sub-50K architecture preserves high-resolution cues across three sports.

02

Coordinate Attention strengthens weak targets while hard-negative supervision suppresses false peaks.

03

At least 0.93 F1 across all three sports with real-time GPU and RK3588 NPU inference.

Method

Network architecture

Three consecutive RGB frames are concatenated into a nine-channel input. A shallow depthwise-separable U-Net limits spatial reduction to one-quarter resolution, Coordinate Attention reweights the bottleneck, and the hard-negative branch is used only during training.

LiteBallNet architecture, Coordinate Attention module, and training-only hard-negative objective
Figure 1. (a) A DS U-Net maps three frames to the current heatmap with 4× spatial reduction. (b) Bottleneck CA applies directional gates gh and gw. (c) The hard-negative mask is training-only.Vector PDF ↗
Input3 × RGB frames

Frames It, It−1, and It−2 are resized to 360 × 640 and concatenated.

Backbone24 / 48 / 96 channels

Two stride-2 stages and full-resolution skip fusion preserve tiny-target evidence.

AttentionBottleneck CA

Axis-specific height and width gates retain position-aware long-range context.

Training onlyλ = 0.08 hard negatives

Structured distractor masks penalize recurrent false peaks without changing inference.

Protocol. The architecture is shared across sports, with one checkpoint trained per sport. A prediction is correct when its distance from the ground-truth centre is within 4 pixels.

Qualitative demos

Cross-sport tracking outputs

Representative project-page clips show the model output on tennis, badminton, and table tennis. Each benchmark uses the same architecture with a separately trained checkpoint.

BenchmarkTennis
Reported F195.17%
Model outputFull-resolution heatmap

Tracking demonstrations. Representative output sequences across the three evaluated sports.

Experiments

Accuracy, capacity, and latency

The main comparison uses 360 × 640 input; dagger-marked FPS values were measured on the same RTX 3060 Laptop.

Average accuracy versus latency across three benchmarks with parameter count represented by bubble area and updated RK3588 latency
Figure 2. Average accuracy versus latency. Bubble area denotes parameters, the star marks LiteBallNet, and the inset reports RK3588 RKNN latency.Vector PDF ↗
MethodParamsFPS†BadmintonTable TennisTennis
Acc.Prec.Rec.F1Acc.Prec.Rec.F1Acc.Prec.Rec.F1
TrackNetV143.2M16.23.8395.9653.8364.8962.7876.9903.7814.8735.7901.9336.8301.8788
TrackNetV243.3M47.03.8644.9672.8657.9136.8640.9786.8729.9228.8124.9596.8356.8933
MonoTrack11.07M30.56.8717.9709.8874.9273.8709.9488.9070.9274.8333.9349.8792.9062
WASB5.7M19.92.8446.9469.8585.9006.8580.9678.8753.9192.8752.9819.8857.9313
MMFNet58.1M14.99.9150.9542.9422.9482.9178.9857.9255.9547.9040.9240.9757.9492
TrackNetV343.3M46.31.9146.9882.9089.9469.9167.9829.9269.9541.8977.9775.9133.9443
TOTNet8.7M6.73.8899.9549.9092.9315.9010.9680.9231.9450.8938.9648.9203.9420
TrackNetV443.3M32.13.9203.9726.9303.9510.9212.9802.9344.9567.9063.9760.9237.9491
BlurBall5.7M17.26.8539.9682.8518.9063.9234.9826.9344.9579.8893.9838.8989.9394
TrackNetV658.0M—.9414.9750.9540.9644.9517.9822.9659.9739.9384.9526.9833.9677
LiteBallNet (Ours)45,64486.30.9267.9602.9129.9361.9849.9794.9924.9859.9544.9494.9540.9517
Table 1. Cross-sport comparison. LiteBallNet is the highest-accuracy model under a 50K-parameter budget.
Badminton0.9361

26 training / 3 test matches

Table Tennis0.9859

5 training / 7 test matches

Tennis0.9517

7 training / 3 test matches

SettingCAHNBadmintonTable TennisTennisAverage
Lightweight baseline——0.90230.95310.92040.9253
+ Coordinate Attention✓—0.93560.98200.94800.9552
+ hard-negative objective—✓0.92340.97320.93940.9453
Full proposed model✓✓0.93610.98590.95170.9579
Table 2. Cross-sport test F1-score ablation under the 4-pixel protocol.
Edge deployment

One inference graph, sport-specific checkpoints

The shared tensor interface allows sport-specific checkpoints to use one RKNN pipeline. Deployment to another supported sport requires checkpoint replacement rather than redesigning the inference graph.

Model size
45,644 parameters
Computation
1.876 GMAC
RTX 3060 Laptop
86.30 FPS
RK3588 NPU · RKNN
37.0 ms · approximately 30 FPS
TrackNetV2 prototype on RK3588
about 1 FPS

Reported values are model-inference measurements; end-to-end camera, decoding, and display latency is not claimed.

RK3588 NPU deployment demonstrations

Mobile recordings show the physical RK3588 prototype in operation, while the corresponding direct outputs provide a clearer view of sport-specific RKNN inference. The overlays report model-inference throughput rather than end-to-end camera-to-display latency.

Mobile-recorded prototype

TennisMobile-recorded RK3588 prototype
Table tennisMobile-recorded RK3588 prototype
BadmintonMobile-recorded RK3588 prototype

Direct RKNN output

TennisRK3588 NPU · RKNN inference
Table tennisRK3588 NPU · RKNN inference
BadmintonRK3588 NPU · RKNN inference
Resources

Paper, code, and contact

Paper

ICASSP 2027 manuscript
The public PDF link will be enabled after the release timing is confirmed.

Pending
Code

LiteBallNet repository
Source code, checkpoints, and deployment resources will be released after acceptance.

After acceptance
Contact

Corresponding authors
liqiang1@sztu.edu.cn · wangpeizheng@sztu.edu.cn

SZTU